This isn’t the first outage that has stopped the nation from working more effectively than any Flemington horse race ever did. This Friday the 13th is just another incident in a long line of failures. Cast your mind back to July 2024, when Microsoft crashed out even worse. Yes, CrowdStrike caught the blame for that one. But everyone with CrowdStrike installed on their work MacBooks was fine, weren’t they?

Same CrowdStrike. Same Friday. The Macs kept working. The Linux boxes kept working. Eight and a half million Windows machines went blue and stayed blue, and Australia copped it first because it was Friday afternoon here. Supermarket checkouts froze. Banks froze. Flights were grounded. CrowdStrike pulled the trigger. Windows was the gun.

The ledger

January this year: Outlook down for close to ten hours. This morning: email, Teams, SharePoint, Copilot and Defender, all at once. Defender is the security product. The guard dog passed out with the rest of the house.

Microsoft’s explanation was that a number of its servers had become “unexpectedly unavailable”. Mate, same. It’s Friday. A few hours in, it admitted its fixes “have not provided the anticipated relief”. I’ve had physios like that.

And for this privilege, on 1 July they put the price up. Between 5 and 43 per cent, depending on your plan. Every year you pay more for a product that works less. Try that in any other trade. Try telling a tradie his new ute costs 40 per cent more and won’t start on Fridays.

Security company, apparently

In 2023, Chinese state hackers stole one of Microsoft’s own signing keys and used it to read the email of senior American officials. The US government’s own review board found the whole thing was preventable and said Microsoft’s security culture needed an overhaul.

Then last year ProPublica revealed that Microsoft had spent the best part of a decade using engineers based in China to maintain the Pentagon’s cloud systems. They were supervised by American “digital escorts” who often lacked the skills to understand what the engineers were typing. The US Defence Secretary called it a breach of trust.

That’s the company whose software runs most of Canberra. Sleep well.

Ask the French

Last year the French Senate put Microsoft France’s top lawyer under oath and asked him one question. Could he guarantee that French citizens’ data would never be handed to the US government without France’s permission? His answer: “No, I cannot guarantee it.”

There’s an American law called the CLOUD Act. If Washington wants the data, Washington gets the data, wherever the server sits. Paris or Sydney, it makes no difference. A data centre in Western Sydney with a kangaroo painted on the side is still an American filing cabinet.

Our response? In 2023 the Prime Minister stood in Washington and celebrated Microsoft spending $5 billion on more of them. We threw a party for our own leash. Labor did the deal, and I don’t remember the Coalition objecting. Neither side has the spine to ask who actually holds the keys to the country.

Europe has worked it out. The German state of Schleswig-Holstein is moving 30,000 public servants off Microsoft and onto Linux and LibreOffice. Denmark’s digital ministry has started down the same road. When German bureaucrats are moving faster than you, have a good hard look at yourself.

How they treat their own

Last year Microsoft sacked 16,000 people while posting record revenue and record profit. In July this year it cut another 4,800. Then, a bit over two weeks ago, 60,000 staff got an email. That’s more than a quarter of the company. Microsoft calls it a “Voluntary Exit Offer”.

Voluntary. When the boss emails a quarter of the building asking who’d like to leave, that’s a sacking with a fruit basket.

And while American workers were being shown the door, Washington approved 5,189 H-1B foreign-worker visas for Microsoft in a single year. Third-highest of any company in America. Microsoft says the visas and the layoffs are “in no way” related. Two US senators, one Republican and one Democrat, wrote to the company demanding answers anyway. When Chuck Grassley and Dick Durbin agree on something, somebody has done something spectacular.

So here’s my question for every Australian business owner. If that’s how Microsoft treats the Americans who built the place, where do you reckon a 12-person accounting firm in Toowoomba sits on its list of priorities?

And then there’s the founder

Bill Gates started having dinners with Jeffrey Epstein in 2011. Epstein had been a convicted sex offender since 2008. That was on the public record, and a man with Gates’s resources can afford a Google search. By his own account the dinners went on for three years.

In January the US Justice Department released three million pages of Epstein files, and there was Gates again. His answer was that he’d been “foolish”. A few weeks later he told his own foundation’s staff he’d had two affairs with Russian women, and insisted neither had anything to do with Epstein’s victims. Melinda has said the Epstein connection was part of why she walked.

No victim has accused Gates of anything and he denies any wrongdoing. Noted. I’m going only on what he admits to, and what he admits to is sitting down to dinner, again and again, with a convicted child sex offender.

The day those files landed, by the way, he was in Melbourne watching the women’s final at the Australian Open. He was in town to ask our government for more money for his foundation’s causes. You couldn’t write it.

Gates left the Microsoft board in 2020. The company he built is the one we’re all still chained to.

The way out

None of this is hard anymore. Macs and Linux on the desks. Email and files hosted in Australia, on open-source software, by a company that answers to Australian law. Automations rebuilt in tools you can pick up and move. AI you choose yourself, including open models running on a box in your own office that answers to nobody in Washington. Either one.

And spare me “but Copilot”. Microsoft has roughly 450 million commercial users. About 15 million of them pay for Copilot. Twenty-nine out of thirty customers looked at Microsoft’s AI and kept their wallets shut.

Every outage, Microsoft tells us it is “aggressively pursuing all pathways”. Good. Here’s one more pathway. The exit. Australia should take it.

Sources

• This morning’s outage: BleepingComputer, HRD
• January outage: The Register
• China-based engineers on Pentagon systems: ProPublica
• French Senate testimony: PPC Land
• Schleswig-Holstein and Denmark: Computing
• Voluntary Exit Offer and the July cuts: Windows Central
• 2025 layoffs and H-1B approvals: Grassley and Durbin letter to Microsoft, Senate Judiciary release
• Microsoft’s reply on visas: Yahoo Finance
• Gates and Epstein: ABC News, Malay Mail / AFP, Deseret News

If we’re building agentic AI systems, when does Apple Silicon make more sense than renting NVIDIA?

This isn’t a fan post. It’s a field note from the operator seat. I don’t care about benchmark chest-thumping. I care about what ships, what it costs, and what breaks at 2am.

The funniest chart in tech

A16Z published a chart (FactSet, Feb 2026) of standardised quarterly capex for the big five:

• Amazon: +42% YOY
• Microsoft: +89% YOY
• Alphabet: +50% YOY
• Meta: +48% YOY
• Apple: −19% YOY

Four companies ploughing $100 billion per quarter into data centres. Apple spending less than last year.

Meanwhile: Mac Minis sold out. Mac Studios on six-week backlog. Someone ran Qwen 3.5 on an iPhone. The M5 Max just shipped with 128GB unified memory running Llama 70B from anywhere.

As @JoshKale put it: “The company spending the least on AI infrastructure accidentally became the AI infrastructure.”

That’s the tension this post unpacks.

Why we care

Our constraints are boring:

• Keep client data private
• Run models without insane infra bills
• Iterate fast
• Keep things stable enough to trust

That lens changes the conversation. Most internet debate is about peak numbers. Most real teams solve for reliability, economics, and speed of execution.

The architecture problem nobody explains well

Most people treat AI infrastructure as a pure compute race. Whoever has the most FLOPS wins. But in inference-heavy workflows — what most businesses actually do — memory is often the real bottleneck.

Traditional GPU setups: CPU and GPU have separate memory pools. Every AI query means data physically moves across a bus. Latency, wasted energy, hard performance ceiling. Architectural, not computational.

Apple’s answer: eliminate the separation entirely. Unified Memory Architecture (UMA) puts CPU, GPU, and Neural Engine on the same memory pool. No data movement. No bus bottleneck.

Practical result: a Mac Studio runs a 70B-parameter model locally, silently, on your desk. Not as fast as an H100 cluster. But you didn’t rent anything, configure anything, or send your data anywhere.

M5 Max: the spec sheet that broke Twitter

Apple just dropped the M5 Max MacBook Pro:

• 18-core CPU with 6 “super cores” — world’s fastest CPU core
• 40-core GPU — rivals an RTX 4070, in a laptop
• 128GB unified memory — more than most servers
• 614 GB/s memory bandwidth — 4x the DGX Spark
• 24-hour battery life
• $3,499

Llama 70B — a model that required a $40,000 GPU cluster 18 months ago — now runs on a laptop at a coffee shop. At 20–30 tok/s, fast enough to actually use.

The local AI revolution just shipped as a consumer product. With a keyboard and a battery.

The numbers that matter

Forget FLOPS. For inference, these metrics determine what you can run and what it costs:

Memory cost per GB:

• Apple M3 Ultra: $18/GB
• NVIDIA DGX Spark: $36/GB
• NVIDIA B200 (DGX): $360/GB

Apple is 20x cheaper per GB than NVIDIA’s best datacentre GPU. Half the price even against the budget-tier Spark.

Energy economics:

• M4 Ultra: ~400 joules per inference task vs cloud GPU: ~10x more
• H100 exceeds 700W sustained; Mac Studio M4 Ultra: a fraction
• A 4-node Mac Studio cluster draws under 250 watts. Whisper-quiet. Under your desk.

Stack Mac Studios and you’re building the cheapest way to run frontier AI models today. NVIDIA has, as one thread put it, “completely missed this segment.”

Doesn’t mean Apple replaces NVIDIA. But if your problem is “fit bigger models locally, keep data onshore, and not torch budget,” the economics are brutal.

Real cluster data (what convinced me)

Jeff Geerling’s 4x Mac Studio cluster testing (Apple-loaned, December 2025) shifted this from “interesting” to “usable.”

Setup: 4x M3 Ultra Mac Studios, 1.5TB total unified memory, Thunderbolt 5. Cost: ~$40,000.

Results:

• DeepSeek V3.1 (671B): 21.1 tok/s single → 27.8 two nodes → 32.5 tok/s four nodes
• Kimi K2 Thinking (1T params): 28 tok/s across the cluster
• Power: Under 250W total. Less than 10W idle per node.
• Geekbench: M3 Ultra beats DGX Spark and AMD AI Max+ 395 in single AND multi-core
• FP64: First small desktop to break 1 Tflop — nearly double the NVIDIA GB10

Geerling’s summary: “A single M3 Ultra Mac Studio has more horsepower than my entire Framework Desktop cluster, using half the power.”

Caveats: RDMA over TB5 is still early. Latency dropped from 300μs (TCP) to 5–9μs (RDMA), but setup requires Recovery Mode and cabling doesn’t scale past 4–7 nodes. No TB5 switches exist.

But “4 quiet boxes under your desk running trillion-parameter models at 28 tok/s for 250 watts” is a different universe from two years ago.

DGX Spark: honest assessment

Not anti-NVIDIA. The Spark has a real use case — it’s a capacity play for models (120B+ in NVFP4) that would crash a 24GB consumer GPU.

But the honest picture:

Carmack’s review (Oct 2025): Power maxing at 100W (not rated 240W). Roughly half quoted performance. Gets “quite hot.” Spontaneous rebooting. Verdict: “My M3 Pro was generating tokens at comparable speeds” for models that fit in 36GB.

Jan 2026 update: 2.5x improvement on prefill and batch. But token generation — what you actually feel — is bandwidth-limited. Physics problem.

Feb 2026: NVIDIA raised the price from $3,999 to $4,699 (18% hike, LPDDR5X supply). The same memory Apple uses. But Apple is the world’s largest LPDDR5X buyer and can ship 512GB Mac Studios while NVIDIA can’t hold pricing on 128GB. Structural supply chain advantage.

The Spark’s 273 GB/s bandwidth looks thin against the M5 Max’s 614 GB/s. AMD’s Strix Halo benchmarks similarly at half the price.

Where Spark shines: Brev hybrid routing, dual-Spark 256GB pools for Llama 405B, and 30+ NIM playbooks. If you need CUDA compatibility, it’s the cheapest entry.

MLX is quietly winning

Apple-side tooling is moving faster than most realise.

A forthcoming paper (vllm-mlx, EuroMLSys ‘26) benchmarked native Apple Silicon inference against llama.cpp:

• 21–87% higher throughput than llama.cpp on Apple Silicon
• M4 Max: up to 525 tok/s on text models (Qwen3-0.6B)
• Continuous batching: 4.3x aggregate throughput at 16 concurrent requests
• Prefix caching: 28x speedup on repeated image queries

Why? MLX exploits UMA properly — lazy evaluation, native quantisation kernels, true zero-copy. llama.cpp was designed for discrete GPUs and adapted; MLX was built for UMA.

Key insight: Apple Silicon’s advantage grows with concurrency. Continuous batching on UMA is fundamentally more efficient because KV cache doesn’t transfer between devices. Single-user tok/s comparisons are misleading — in multi-user serving, the UMA advantage compounds.

Demand signals match: Mac Minis backordered, Mac Studios six-week wait. When Alibaba dropped Qwen 3.5, MLX support landed same-day — running on an iPhone within hours.

I track a simple metric: TTLD (Time To Local Deployment) — model release to running privately. For agentic workflows where privacy and iteration speed beat peak throughput, Apple + MLX consistently wins TTLD even when it loses raw numbers.

Why DeepSeek V4 changes the hardware calculus

This is the part nobody’s connecting yet.

Standard transformers waste expensive GPU compute on two fundamentally different tasks:

1. Static Recall — “What’s the syntax for a Python list comprehension?” Memory lookups. No reasoning needed.
2. Dynamic Reasoning — Logic, composition, novel problem-solving. Needs full compute.

Every model today uses the same expensive hardware for both. Like hiring a surgeon to do your filing.

DeepSeek V4’s Engram architecture separates them. Static knowledge offloads to an O(1) hash-based lookup table in system DRAM — 100B parameters in regular memory, not GPU memory. Throughput penalty: less than 3%.

The hardware implication: high-bandwidth system memory is now as valuable as GPU FLOPS.

On traditional x86+GPU, Engram lookups go through the PCIe bus — bottleneck. On Apple Silicon’s UMA, they’re in the same memory pool as GPU compute. Zero-cost access.

V4 activates only 32B parameters from its 1 trillion total (fewer than V3 despite being 50% larger). Less GPU compute needed, more memory needed. Apple’s $/GB advantage dominates exactly the cost structure V4 optimises for.

The convergence thesis: As architectures evolve to separate memory from reasoning — Engram being the first — Apple Silicon’s unified memory becomes more advantageous. The trajectory:

• 2024: “Apple can’t do AI” (the training narrative)
• 2025: “Apple Silicon is interesting for inference” (the memory narrative)
• 2026: “Apple Silicon + Engram-style models = optimal local inference” (convergence)

As more models adopt conditional memory and knowledge offloading, the gap widens.

ANE: watch this space

The Neural Engine deserves a mention — not because it’s production-ready for LLMs, but because it signals untapped headroom.

M4’s Neural Engine: 38 TOPS. M5 puts neural accelerators inside each GPU core — generational architecture shift. A reverse-engineering project (maderix/ANE) has demonstrated training and backprop on the ANE with compelling microbenchmarks.

But today: private APIs, fragility risk, CPU fallbacks, potential breakage every macOS update. Research territory, not production dependency. File under “strategic headroom.”

The honest counterarguments

Throughput at scale: Need 100+ tok/s on the largest open models? Apple loses. Two Mac Studios running Kimi K2.5 at 4-bit quant: 10–12 tok/s. Two AMD TBv2 cards might be cheaper for raw throughput. Apple wins $/GB but can lose throughput-per-dollar at the high end.

RDMA is green: TB5 RDMA is a strategic signal, not production-grade clustering. CPU spikes (900%+) from TB Bridge loops. Recovery Mode access required. No TB5 switches. Full mesh means N-1 cables per node. Beyond 4–7 nodes: unsolved.

Cluster management pain: Exo clustering works and auto-discovers, but it’s not Kubernetes. Don’t expect enterprise ops maturity.

Training is still NVIDIA: Heavy training and fine-tuning remain NVIDIA territory. CUDA ecosystem, mature distributed training, raw scale. Training-heavy workload? This conversation doesn’t apply.

This isn’t a religion. Any real strategy here is mixed. The question isn’t Apple or NVIDIA — it’s knowing when each wins.

How we’re deciding

Apple-first when:

• Inference-heavy workload
• Privacy or data locality matters
• Memory-per-dollar matters
• Fast local iteration is valuable
• Data stays onshore without cloud bills

NVIDIA/cloud-first when:

• High throughput target (100+ tok/s on large models)
• Training or fine-tuning heavy
• Enterprise ops maturity required
• CUDA ecosystem compatibility needed

In practice: increasingly hybrid. Apple for local core paths and private data. Cloud for peak compute when needed.

Personal Computing v2

Karpathy framed it on Latent Space: “As we leave the cloud for Personal/Private AI, some signs of Personal Computing v2 are being born in Exolabs and Apple MLX work.”

Two eras:

1. PC v1 (1980s): Computation moved from mainframes to desktops.
2. PC v2 (2026+): AI inference moves from data centres to desktops.

Apple won the first transition. They’re building infrastructure for the second — not with data centres, but with architecture.

The data sovereignty angle isn’t just privacy idealism. It’s economic inevitability. As models grow and API costs compound, the break-even for local inference keeps moving earlier. A Mac Mini M4 Pro pays for itself vs cloud H100 rental ($2.39/hr) in roughly 1,000 hours — about 6 weeks of continuous inference.

Where I’ve landed

A year ago, serious local AI on Apple felt niche. Now it feels like a legitimate operating mode.

Not because Apple “won AI.” Because architecture, economics, and tooling velocity combined into something too practical to ignore.

Look at that capex chart again. Big Tech is spending hundreds of billions on top-down AI infrastructure — massive data centres, custom chips, power plants. Apple spent less and ended up with sold-out hardware that developers are clustering to run trillion-parameter models under their desks.

Nobody planned that. It’s what happens when you build the right architecture and demand finds you.

If you’re building agentic systems and haven’t pressure-tested local inference, it’s worth doing. The numbers might surprise you.

Next up: concrete reference builds (Mac Mini / Mac Studio tiers), expected model envelopes, and where each setup starts to fall over.

This is for small business owners — 5 to 50 employees — drowning in admin and wondering if AI can actually help. Short answer: yes. But not the way most people think.

What’s an AI Agent (In Plain English)?

A chatbot answers questions. An agent does work.

Asking someone for directions vs. hiring a driver. A chatbot gives you the route. An agent drives you there, finds parking, and texts your friend that you’ve arrived.

In business terms, an AI agent can:

• Read an email, understand the request, pull the client file, draft a response, and send it — untouched
• Answer your phone after hours, qualify urgency, book an appointment if needed, and send you a morning summary
• Process an invoice, match it to a PO, flag discrepancies, route for approval

Not science fiction. Businesses are deploying this right now.

The 7 Workflows Worth Automating

Ranked by time saved per dollar spent. Start at the top.

1. Appointment Scheduling & Reminders

Time saved: 8–15 hours/week for a busy practice Best for: Health practitioners, consultants, trades with bookings

Customer wants to book → calls/emails → staff checks availability → books → confirms → reminds → handles reschedules. Every step can be handled by an AI agent.

Real numbers: A chiro seeing 200+ patients/week spent 20 hours/week on scheduling admin. AI cut that to 3 hours of exception handling.

2. Client Follow-Up & Recall

Time saved: 5–10 hours/week Best for: Any business with repeat customers

Businesses lose thousands monthly from customers who simply forget to come back. Your dentist didn’t forget about you — they didn’t have a system that remembered.

An AI agent monitors your client database, identifies overdue contacts, sends personalised follow-ups via their preferred channel, tracks responses, and escalates only when needed.

3. Invoice Processing & Accounts Payable

Time saved: 5–12 hours/week Best for: Anyone processing 50+ invoices/month

Old way: Invoice arrives → someone opens it → types details into accounting software → matches to PO → gets approval → schedules payment.

AI way: Invoice arrives → agent extracts details (OCR + understanding) → auto-matches PO → flags discrepancies → routes for approval → schedules payment → updates books.

Firms processing invoices for 40+ clients routinely report 60–70% time reduction. Staff shift from data entry to quality control.

4. Customer Inquiry Triage

Time saved: 3–8 hours/week Best for: High-volume businesses (trades, accommodation, services)

Not every inquiry deserves the same response time. A luxury accommodation getting 50 booking inquiries daily needs to prioritise the $5,000 week-long stay over the “do you allow pets?” question.

AI reads every inquiry, classifies urgency and value, drafts appropriate responses, routes high-value ones to a human immediately. Low-value inquiries get helpful, personalised auto-responses.

5. Document Intake & Processing

Time saved: 4–10 hours/week Best for: Professional services (accounting, legal, financial planning)

Clients send documents in every format imaginable — receipt photos, bank statement PDFs, scanned handwritten notes. Someone spends hours sorting, naming, filing, and extracting the relevant info.

AI handles the full intake: classify, extract key data, file correctly, update the client record, flag anything unusual.

6. Review Management & Reputation

Time saved: 2–4 hours/week Best for: Any customer-facing business

AI monitors reviews across Google, Facebook, and industry platforms. Drafts responses (professional for negative, grateful for positive), flags concerning patterns, and prompts happy customers to leave reviews at the right moment.

7. Supplier & Vendor Communication

Time saved: 3–6 hours/week Best for: Events, construction, any multi-supplier business

Requesting quotes, confirming orders, chasing deliveries, updating timelines — all repetitive. AI handles routine communication, tracks responses, involves you only when decisions are needed.

What This Costs

Let’s be direct.

Hiring an admin/VA: $25–35/hour × 20 hours/week = $2,000–2,800/month.

AI agents: $500–2,000/month + one-time setup of $2,000–5,000.

Typical payback: 3–6 months, then pure savings. Unlike a person, agents work 24/7, don’t take sick days, and handle volume spikes without overtime.

Run the numbers for your situation: AI ROI Calculator.

How to Start Without Blowing Your Budget

1. Pick ONE workflow — the biggest time sink
2. Measure current cost — hours × hourly rate of whoever’s doing it
3. Pilot for 30–60 days — one workflow, measurable outcomes
4. Measure the result — hours saved, errors reduced, revenue recovered
5. Scale or stop — works? Add another. Doesn’t? You risked one month.

Worst approach: automate everything at once. Best approach: prove ROI on one workflow, then expand.

Objections, Answered Honestly

“My business is too unique.” It’s probably not. These workflows are universal. Implementation varies; patterns don’t.

“Customers want a real person.” Some do. Most want a fast, accurate response. AI handling the first 80% means your team has more time for the 20% that genuinely needs a human.

“What about mistakes?” AI makes mistakes. So do humans. The difference: AI mistakes are consistent (and fixable). Human mistakes are random. Set confidence thresholds — anything below goes to a human.

“We tried a chatbot and it was terrible.” Chatbots in 2023 were terrible. AI agents in 2026 are a different technology entirely. Calculator vs. spreadsheet.

Next Step

If you’re a small business on the Sunshine Coast (or anywhere in Queensland) spending 10+ hours/week on admin that doesn’t directly make money, we should talk.

Free 15-minute discovery call — no pitch, just a conversation about where AI saves you the most time. Fit? We’ll map a pilot. No fit? You’ll walk away with useful ideas.

Book a Discovery Call →

Richard Dobson is the founder of Dobson Development, helping Queensland businesses automate operations with AI agents. Based on the Sunshine Coast.

There’s a lot of marketing around “agentic AI” in security right now and not much clarity about what’s real. I’ve spent months digging into the technical architectures behind the platforms making serious claims. Here’s what I found.

Definitions Worth Getting Right

These terms get used interchangeably. They shouldn’t be.

• Copilot — AI that helps a human work faster (answers questions, drafts reports, suggests next steps)
• Automation — Deterministic if/then logic at machine speed (traditional SOAR)
• Agentic AI — Systems that reason about goals, plan action sequences, use tools autonomously, and adapt when things break

The third category is genuinely new. The first two have existed for years. The distinction matters.

The Incumbents

CrowdStrike — Charlotte AI & AgentWorks

Charlotte AI is more technically interesting than it gets credit for.

Core claim: Charlotte was trained on elite analyst judgment, not just threat data. When an alert arrives, it runs a model that learned what a skilled analyst with deep context would do — handling novel patterns no playbook covers.

Three layers:

1. Detection Triage — Ingests Falcon telemetry, autonomously triages, filters false positives, surfaces what matters. Continuously retrains as real analysts validate or override.

2. Charlotte Agentic SOAR — Structured SOAR logic plus agentic reasoning. Agents think and decide when situations don’t match playbooks. Analysts can inject context mid-investigation.

3. AgentWorks — No-code platform for building custom security agents in natural language. Define goal, data sources, guardrails — no engineering required.

Strength: Falcon telemetry. Nobody else has that volume of endpoint data to train on.

Limitation: Deep Falcon ecosystem lock-in. Less valuable if you’re not CrowdStrike-heavy.

Palo Alto Networks — Cortex XSIAM + AgentiX

XSIAM is the most ambitious platform consolidation bet in enterprise SOC. AgentiX is the agentic layer on top.

• XDL (Extended Data Layer) — Unified data fabric normalising endpoint, network, identity, cloud, and third-party telemetry into a single semantic layer.
• 2,600+ ML detection models running continuously, with 100% MITRE ATT&CK coverage in their latest evaluation.
• AgentiX — AI agent workforce that plans, reasons, and acts autonomously. Specialised agents with defined roles and guardrails.

Published numbers: 98% MTTR reduction, 99% alert noise reduction, 300% ROI. Oneida Nation claims 43-second incident resolution. Vendor metrics deserve scepticism, but independent MITRE ATT&CK results are harder to wave away.

Strength: Platform consolidation. Once on Cortex, switching costs are enormous.

Limitation: Enterprise pricing. Not targeting SMB.

The Challengers

Torq — HyperSOC & Multi-Agent Architecture

Torq makes the most interesting architectural argument among challengers: autonomous SOC should be a Multi-Agent System (MAS), not a single AI model.

HyperSOC-2o orchestrates specialised agents across triage, investigation, response, and post-incident summary — each optimised for its function. Deterministic automation handles known workflows reliably; the agentic layer handles ambiguity.

Carvana’s result: Torq handles 100% of Tier-1 alerts and automated 41 runbooks within one month. The one-month timeframe matters — traditional SOAR deployments took quarters.

Strength: MSSP/MDR focus. Built for providers running SOC at scale.

Limitation: Middleware play. Depends on your existing tools working well underneath.

Tuskira — The AI Defence Mesh

Tuskira is the most technically differentiated startup here and probably the least understood.

They’re building an AI Defence Mesh — architecturally distinct from everything else. Five steps:

1. Ingest and Normalise: 150+ integrations across SIEM, EDR, CSPM, IAM, WAF, GRC, and cloud configs.

2. Build the Digital Twin: Continuously updated model of your environment — cloud topology, network reachability, identity relationships, control coverage. Not a static diagram; a live model of how attackers could move today.

3. AI Simulation & Validation: Attack paths continuously tested against your defences. Simulates real attacks against the digital twin to identify what’s truly exploitable — filtering false positives before humans ever see them.

4. AI Analysts Act: Role-based AI analysts (Vulnerability, Zero-Day, Threat Advisory, Remediation) triage with full digital twin context, correlate across tools, tune defences automatically.

5. Closed-Loop Feedback: Every action, simulation result, and posture drift feeds back into the mesh.

Published outcomes: 98% alert noise reduction, 60% fewer attack paths, 5-minute triage.

This is preemptive defence, not reactive triage. The core innovation: digital twin plus continuous simulation — finding what’s exploitable before attackers do.

Strength: Only platform doing live attack simulation at this scale as a continuous background process.

Limitation: Early stage. Enterprise-complexity deployment.

The Triage Specialists

Radiant Security — Transparent Agentic Triage

Radiant solves a problem most platforms handle poorly: explainability.

Every escalation or dismissal includes full traceability — which sources were queried, what patterns detected, why that conclusion. Analysts validate reasoning, not just accept it.

Smart economic bet too: built-in security data lake at up to 85% cost reduction vs traditional SIEM, flat-rate pricing.

Strength: Transparent reasoning for regulated industries where “the AI said so” doesn’t fly.

Prophet Security — Speed at Scale

Prophet’s numbers are the sharpest in category: alerts investigated in under 3 minutes versus the 30-minute human baseline — 90%+ MTTI reduction.

Strong data privacy argument: single-tenant architecture, no customer data used to train models. In an industry deeply sceptical of AI handling sensitive data, that’s a meaningful trust signal.

Unusually specific CISO testimonials — Instacart, Clari, Spotnana, Zip — describing concrete outcomes, not generic endorsements.

Dropzone AI — Autonomous Tier-1

The most narrowly focused pure-play and arguably the most practically deployable. One job: autonomous Tier-1 alert investigation.

Replicates the investigative sequences a senior analyst would follow — pulling data, forming hypotheses, testing them, writing findings. Deploys in minutes via API keys, not months of professional services. Learns your environment continuously.

Before/after: MTTR from hours to minutes, manual analysis from 25 to 2 minutes per alert, investigation coverage from 30% to 100%.

Strength: Easiest deployment in the category. Day-1 value. Ideal when you can’t afford a six-month platform rollout.

Exabeam Nova — UEBA-Native Agent

Nova leverages something unique: a decade of UEBA data and models.

Where other platforms start from scratch with agent reasoning, Nova has deep behavioural baselines. Context isn’t just “this IP hit a suspicious URL” — it’s “this user’s behaviour is 6.2 standard deviations from their 90-day baseline, they’ve never accessed this system, and three peer-group users showed the same pattern last week.”

Their new AI Agent Behaviour Analytics extends this to monitoring deployed AI agents themselves — detecting anomalous agent behaviour. First to market on that.

What It All Adds Up To

The architectural divergence is real. Tuskira solves a fundamentally different problem than Dropzone. Charlotte AI does something different from Radiant. These platforms have genuinely different theories of what “autonomous SOC” means.

Deployment complexity still matters. The most sophisticated platforms (Tuskira, XSIAM) are the hardest to deploy. Need value this quarter? Dropzone and Prophet offer a faster path.

Data moats will decide long-term winners. CrowdStrike has Falcon telemetry. Palo Alto has firewall + XDR + cloud data. Exabeam has 10 years of UEBA baselines. Pure-plays bet that reasoning quality beats data volume. The incumbents will catch up on reasoning while keeping their data advantage.

“Human-in-the-loop” means wildly different things. Ranges from “human approves every action” to “human reads the weekly summary.” Worth understanding exactly what you’re buying.

Tier-1 triage is a solved problem. If your SOC still manually triages every alert, that’s table-stakes automation now. The interesting question is Tier 2 and 3 — real investigation and response, not just sorting.

The One Question Worth Asking Every Vendor

“Show me a real investigation, end to end, on a real alert, with no staging. What did the agent do, in what sequence, and how did it know to do that?”

The ones who answer clearly are building real products. The ones who pivot to a slide deck aren’t there yet.

Evaluating how agentic AI fits your SOC roadmap? Get in touch — we help businesses navigate this space.

Every security vendor now claims to be “AI-powered.” Most of them bolted a chatbot onto their existing product and updated their website. A few are genuinely transforming how security operations work.

This guide helps you tell the difference.

What AI Is Actually Doing in Security (Right Now, in Production)

Skip the breathless predictions. Here’s what’s working today.

Alert triage that doesn’t suck

This is AI’s biggest practical win. Modern SOCs generate thousands of alerts daily. AI trained on historical alert data and analyst decisions can now:

• Classify alerts at 85–95% accuracy, matching experienced analysts
• Correlate related alerts into incidents automatically
• Prioritise by business context — a domain controller alert gets treated differently than the same alert on a test box
• Suppress known false positives without creating blind spots

Microsoft Sentinel’s fusion detection, CrowdStrike’s Charlotte AI, and SentinelOne’s Purple AI lead here. The result: analysts investigate real threats instead of triaging noise.

Threat detection beyond signatures

Traditional detection only catches what you’ve written a rule for. AI adds:

• Behavioural analysis — detecting anomalous user and entity behaviour without predefined rules
• Novel threat detection — identifying attack patterns that don’t match known signatures
• Living-off-the-land detection — spotting malicious use of legitimate tools (PowerShell, WMI) by context, not command matching

Genuinely useful, with a caveat: AI detection generates its own false positives. The best tools learn from analyst feedback over time.

Investigation on autopilot

AI copilots now:

• Summarise incidents in plain English, pulling together logs, alerts, and context
• Suggest response actions based on playbooks and analyst history
• Execute containment — isolating hosts, disabling accounts, blocking IPs — with human approval or fully automated for high-confidence scenarios
• Generate investigation queries — translating English into KQL, SPL, or whatever your SIEM speaks

This isn’t replacing analysts. It’s removing the tedious parts. A junior analyst with an AI copilot can now perform initial triage that previously required years of experience.

Vulnerability prioritisation that’s actually useful

AI tools predict which vulnerabilities will actually be exploited, factoring in exploit availability, your specific environment, threat actor activity, and asset criticality.

Turns a list of 10,000 vulnerabilities into the 200 that actually matter.

The Categories Worth Knowing

AI-powered SIEM and SOAR

What: Log ingestion, threat detection, automated response — AI enhancing every stage.

Key players: Microsoft Sentinel + Copilot for Security, Splunk AI Assistant, Google Chronicle + Gemini, Elastic AI Assistant.

Where AI helps most: Query generation, alert correlation, investigation summaries, playbook recommendations.

Our take: If you’re choosing a SIEM in 2026, AI capabilities should be a key criterion. The productivity gap between AI-enhanced and traditional is significant. If you’re figuring out what to feed it, our Log Source Priority Calculator helps prioritise by detection value and cost.

AI-native SOC platforms

What: Purpose-built platforms using AI as the core engine, not an add-on.

Key players: Torq Hyperautomation, Swimlane Turbine, Intezer, Dropzone AI.

Where AI helps most: Autonomous alert triage, investigation, tier-1 augmentation.

Our take: Most valuable for organisations that can’t staff a 24/7 SOC. AI handling tier-1 triage with humans on escalations is a practical SMB model. See our SOC Automation Playbook Library for pre-built workflows.

AI for email security

What: Phishing, BEC, and social engineering detection using NLP and behavioural analysis.

Key players: Abnormal Security, Material Security, Proofpoint (with AI), Tessian (now Proofpoint).

Where AI helps most: Catching sophisticated phishing that bypasses traditional filters — especially BEC attacks with no malicious links or attachments.

Our take: One of the most mature AI security categories. If you’re still relying solely on Microsoft’s built-in email filtering, an AI email layer is one of the highest-ROI security investments you can make.

AI-powered identity security

What: Detecting identity-based attacks — compromised credentials, privilege escalation, lateral movement.

Key players: Microsoft Entra ID Protection, CrowdStrike Identity Threat Protection, SentinelOne Identity, Silverfort.

Where AI helps most: Impossible travel, unusual access patterns, privilege escalation, service account abuse.

Our take: Identity is the new perimeter. If Active Directory or Entra ID is central to your environment, this is high-value.

AI for code security

What: Finding vulnerabilities in code, suggesting fixes, detecting insecure patterns during development.

Key players: Snyk (DeepCode AI), GitHub Copilot Autofix, Semgrep AI, Checkmarx AI.

Our take: Useful if you have active dev teams. The auto-fix capabilities genuinely save time.

AI security for AI

What: Securing your organisation’s AI usage — LLM input/output protection, prompt injection detection, data leakage prevention.

Key players: Protect AI, Robust Intelligence, Lakera, CalypsoAI.

Our take: If your employees use LLMs (they almost certainly do, whether you know it or not), this category matters. Our Upload-to-LLM Risk Scanner can help assess exposure.

How to Evaluate AI Security Tools (Without Getting Sold)

1. “What specific problem does the AI solve?”

If the answer is vague — “it uses AI to improve security” — walk away. Good tools solve measurable problems: alert volume down X%, MTTR improved by Y minutes, Z% more true positives.

2. Demand a POC in YOUR environment

AI models perform differently on different data. A brilliant demo may struggle with your log sources and infrastructure. Always run a 30-day POC.

3. Check the feedback loop

Does the model learn from your analysts’ decisions? How quickly? If dismissing a false positive doesn’t improve future accuracy, the AI is static — and static AI degrades.

4. Understand the data appetite

What log sources does it need? How much historical data before it’s accurate? Do you get value on day one?

5. Evaluate the human-AI handoff

AI should augment your team, not create busywork. Look for clear decision explanations (not just confidence scores), easy override mechanisms, and integration with your existing tools.

6. Push on false positive rates

Every vendor claims low rates. Ask for data from organisations your size and industry. Better yet: measure it yourself during the POC.

🤖 Explore 50+ AI Security Tools — Free Tool

Our AI Security Landscape Explorer maps 50+ AI security tools across every category — SIEM, SOC, email, identity, AppSec, and more. Filter by use case, company size, and budget to find the right tools for your stack.

Explore the AI security landscape →

Where AI Security Still Falls Short

Strategy is still a human job

AI is excellent at operational efficiency. It’s terrible at strategic decisions — which risks to accept, how to allocate budget, what your programme should prioritise. That still requires judgment.

Novel attacks remain hard

AI detection is pattern-based. Truly novel techniques may evade detection until models retrain. Defence in depth still matters.

The “AI wrapper” problem

Many “AI-powered” tools are thin wrappers around OpenAI or Anthropic APIs. They’ll summarise your alerts using GPT — that’s not the same as purpose-built security AI trained on security data. Ask whether their AI is proprietary, fine-tuned, or just an API call.

Data sovereignty

For Australian organisations: Where is the data processed? Is it used to train models for other customers? Does it comply with the Privacy Act 1988?

Building an AI-Enhanced Security Stack

For most Australian SMBs, the practical path:

1. Start with your SIEM/XDR. AI capabilities in Sentinel, CrowdStrike, or SentinelOne are your quickest path to value. You’re probably already paying for them.

2. Add AI email security. Clear ROI, mature category. Abnormal Security or Proofpoint with AI are solid.

3. Build automation. Use our AI Prompt Library for Security Teams for effective security AI prompts.

4. Fill gaps with specialist tools. Our Security Control Coverage Calculator identifies where AI tools could help.

The Bottom Line

AI is the most significant shift in security ops since the SIEM was invented. But the value isn’t in the AI itself — it’s in solving real problems: too many alerts, too few analysts, too much complexity.

Focus on tools that solve YOUR problems. Run POCs in YOUR environment. Don’t let marketing buzzwords substitute for evaluation.

Explore 50+ AI security tools in our free landscape explorer →

Every EDR vendor’s website says “Contact sales for pricing.” Translation: you’re walking into a negotiation blind, with no idea whether the quote is competitive or highway robbery.

We’ve reviewed dozens of EDR quotes across Australian businesses from 20 to 2,000 endpoints. This guide shares what companies are actually paying — not list prices, not “starting from” marketing fiction.

The Shortlist (2026)

The EDR market has consolidated. For most Australian SMBs, the realistic options:

• CrowdStrike Falcon — Market leader. Premium pricing, strong detection, cloud-native.
• SentinelOne Singularity — The main CrowdStrike alternative. Competitive features, often cheaper.
• Microsoft Defender for Endpoint — Bundled with M365 E5. “Free” if you’re already paying.
• Sophos Intercept X — Popular with MSPs. Strong value for smaller orgs.
• Trend Micro Vision One — Solid if you’re already in the Trend ecosystem.
• Palo Alto Cortex XDR — Enterprise-focused. Less common in SMB.

Others exist (Cybereason, Trellix, Elastic), but these six dominate Australian deployments.

What You’ll Actually Pay (AUD, per endpoint/month, annual contracts)

CrowdStrike Falcon

In practice: Most SMBs land on Pro or Enterprise at $14–$22 after negotiation. CrowdStrike’s list prices are high, but they negotiate — especially for 100+ endpoints or multi-year deals.

Watch out for: Module creep. CrowdStrike’s base product is strong, but they’ll push add-ons for identity, cloud, and log management at $3–$8/endpoint/month each. Decide what you need before the sales call.

SentinelOne Singularity

In practice: 15–25% cheaper than CrowdStrike for equivalent functionality. Most SMBs land on Control or Complete at $11–$18.

Watch out for: Data retention tiers. Base retention (14 days) is often insufficient for compliance. Extended retention (90–365 days) is an upsell. Factor this in.

Microsoft Defender for Endpoint

In practice: If you’re on M365 E5, start here — you’re already paying. Defender for Business is surprisingly capable for under 300 users.

Watch out for: Per-user vs per-endpoint licensing. Multiple devices per user? Microsoft’s model can be cheaper. Shared devices, kiosks, or servers? Gets complicated. Servers need separate licences.

Sophos Intercept X

In practice: The value play, especially via MSP partners. Their MDR tier is worth considering if you lack in-house security — often cheaper than buying EDR + separate MDR.

Watch out for: MSP pricing varies significantly by provider margin. Get multiple Sophos partner quotes.

The Hidden Costs Nobody Mentions

Implementation and tuning

Budget 20–40 hours for initial deployment and policy tuning (100-endpoint environment). MSP setup fee: typically $3,000–$8,000.

Ongoing management

EDR isn’t set-and-forget. Someone reviews alerts, tunes policies, responds to detections. Options:

• In-house: 5–10 hours/week of IT time
• MSP/MDR: $3–$10/endpoint/month on top of the licence

Alert fatigue tax

A poorly tuned EDR generates hundreds of alerts. Each costs investigation time. If you don’t invest in tuning, you pay in wasted hours and missed real threats.

Integration costs

Connecting EDR to SIEM, SOAR, or ticketing adds complexity. Some vendors charge extra for API access or premium integrations.

Worth Paying For (And What Isn’t)

Worth it

• MDR — If you don’t have dedicated security staff, this is the single most valuable add-on. A $15/endpoint EDR+MDR beats a $25/endpoint premium EDR nobody watches.
• Extended data retention — Compliance requirements or incident investigation? Pay for at least 90 days.
• Identity protection — If Active Directory is central to your environment, genuinely useful.

Often not worth it

• Cloud security modules — Unless you’re running significant IaaS workloads, your cloud provider’s native tools suffice.
• Vulnerability management add-ons — Dedicated tools (Tenable, Qualys) are typically better and cheaper.
• Premium threat intel feeds — Unless you have a threat intel analyst on staff, you won’t use it. Base intel is fine for SMBs.

Negotiation Tips That Work

1. Get three quotes minimum. Shortlist CrowdStrike, SentinelOne, and one other. Let each know they’re competing. This alone knocks 15–20% off.

2. Ask for “competitive displacement” pricing. Replacing existing AV or EDR? Every major vendor has a switch programme — routinely 20–30% off.

3. Multi-year for discount (read the terms). Three-year deals save 15–25%, but ensure exit clauses and locked pricing.

4. Negotiate on retention, not seats. If per-endpoint price is firm, push for extended retention, extra modules, or pro services credits.

5. Time your purchase. Vendor fiscal year-ends (CrowdStrike/SentinelOne: January, Microsoft: June) = maximum sales motivation.

6. Don’t buy what you don’t need. Vendors bundle to inflate deals. If you don’t need USB control or firewall management, don’t pay for the tier that includes them.

💰 Check If You’re Overpaying — Free Tool

Our Cyber Vendor Pricing Reality Checker compares what you’re paying against real-world benchmarks for your size and region.

Check your pricing →

The Honest Recommendation

Already on M365 E5? Start with Defender. It’s included, it’s improved massively, and it integrates natively. Layer on dedicated EDR later if needed.

Want best-of-breed with budget? CrowdStrike Falcon Pro or Enterprise. Market leader for a reason. But skip Elite unless you genuinely need identity protection and full XDR.

Strong protection, lower price? SentinelOne Complete. Matches CrowdStrike on most independent evaluations, 15–25% cheaper.

Small business (<100 endpoints), cost-sensitive? Sophos Intercept X via MSP, or Defender for Business. Both capable and affordable.

No security staff? Prioritise MDR over features. Sophos MDR, CrowdStrike Complete, or SentinelOne Vigilance will serve you better than a premium console nobody monitors.

The Bigger Picture

EDR is one piece. Before optimising your EDR spend, understand your overall posture:

• Security Control Coverage Calculator — map where your tools cover (and miss) common techniques
• Security Stack Maturity Score — see how your programme compares
• AI Security Landscape Explorer — 50+ tools across every category

The Bottom Line

EDR pricing is opaque by design — vendors profit from information asymmetry. Arm yourself with real data before you negotiate.

The most expensive EDR isn’t the best one. A well-tuned, actively monitored mid-tier EDR will outperform a premium product that’s deployed and forgotten. Every time.

Check your EDR pricing against real-world benchmarks →

If you run IT for an Australian business, the Essential Eight isn’t optional anymore — it’s table stakes. Government contracts, cyber insurance, not being the next breach headline: all roads lead here.

This guide cuts the jargon. What the E8 actually requires, how maturity levels work in practice, and a realistic path to compliance — even if your “security team” is you and an MSP.

The Essential Eight, Briefly

Eight mitigation strategies from the Australian Signals Directorate (ASD), singled out as the most effective baseline controls:

1. Application control — only approved apps execute
2. Patch applications — browsers, PDF readers, Office: patched promptly
3. Configure Office macro settings — block macros from the internet
4. User application hardening — disable unnecessary features (Flash, ads, Java)
5. Restrict admin privileges — limit who has admin and what they can do
6. Patch operating systems — keep current, retire unsupported
7. Multi-factor authentication — on remote access, privileged actions, sensitive data
8. Regular backups — maintain and test

Nothing exotic. That’s the point. These address the most common attack vectors — phishing, unpatched vulns, credential theft, ransomware — with proven, practical controls.

Why It Matters More Than Ever

Government contracts: The updated PSPF mandates at least Maturity Level Two for agencies. Sell to government — directly or as a sub — and you’ll be asked to demonstrate your E8 posture.

Insurance: Underwriters aren’t just asking “do you have MFA?” anymore. Many now reference the Essential Eight explicitly. Your maturity level affects premiums and whether you get coverage at all.

The threat landscape: Australia saw a 23% increase in reported cyber incidents in 2025. SMBs were disproportionately hit. Ransomware groups specifically target organisations they perceive as soft.

De facto standard: Even without a private sector mandate, regulators, industry bodies, and boards increasingly reference the E8. Get breached without basic ASD-recommended controls in place? Expect hard questions.

Maturity Levels — What They Actually Mean

Level Zero

You haven’t meaningfully implemented the strategy. Running Windows endpoints without application control? Admin accounts without MFA? You’re here.

Level One

Basic implementation. Most common attack vectors addressed, but gaps against sophisticated threats.

• Patching internet-facing apps within two weeks
• MFA on remote access (maybe not all privileged access)
• Backups exist but aren’t tested regularly

Level Two

Strong implementation — what the Australian Government targets for non-critical systems.

• Critical vulns patched within 48 hours
• Phishing-resistant MFA (not SMS)
• Application control on all workstations and servers
• Admin privileges tightly scoped with separate accounts
• Backups tested, stored offline or immutable

Level Three

Highest level. Designed for nation-state-level adversaries. Most SMBs don’t need this.

• Real-time application control with continuous monitoring
• Automated patching with near-zero delay
• Privileged access workstations for all admin tasks
• Hardware tokens and conditional access

Where Most Australian SMBs Actually Stand

Let’s be honest:

MFA is usually strongest because cloud providers pushed it hard. Application control and user application hardening are almost always weakest — they require tooling and policy most SMBs haven’t invested in.

A Practical Path to Level Two

You don’t need to do everything at once.

Phase 1: Quick wins (Weeks 1–4)

MFA everywhere. M365 or Google Workspace? Enforce MFA for all users with Conditional Access. Move admins to phishing-resistant methods (passkeys, FIDO2 keys).

Block macros. Configure Group Policy to block macros in files downloaded from the internet. Policy change, not product purchase.

Audit admin privileges. Find every account with admin rights. Remove admin from daily-use accounts. Create separate admin accounts for IT staff.

Phase 2: Foundations (Months 2–3)

Automate patching. Intune, WSUS, or Automox. Set a 48-hour SLA for critical patches.

Harden user applications. Kill Flash (if somehow still alive), block browser ads via policy, disable unnecessary extensions, block Java and unverified downloads.

Harden backups. 3-2-1 rule. Test restores quarterly. Implement immutable or air-gapped copies against ransomware.

Phase 3: The hard stuff (Months 3–6)

Application control. The most operationally complex strategy. Start with WDAC or AppLocker on a pilot group. Build your approved app baseline. Expect user friction — communicate early and often.

Continuous improvement. Log and monitor to verify controls work. Review and tighten quarterly.

Mistakes That Will Bite You

Treating it as a checkbox exercise. If your application control policy exists on paper but isn’t enforced technically, you’re Level Zero. Documentation without enforcement is decoration.

Skipping the hard strategies. Application control and user application hardening are consistently weakest because they’re hard. They’re in the E8 precisely because they’re effective. Don’t dodge them.

Over-scoping. Aim for Level Two, not Three. L3 is for high-value targets facing APTs. For most SMBs, L2 provides excellent protection without the operational overhead.

Forgetting cloud. The E8 applies to SaaS, cloud servers, and remote endpoints too. Everything in scope.

Not testing backups. An untested backup is a backup that doesn’t exist. Quarterly restore tests. Documented results. No exceptions.

Assessing Where You Stand

Before improving, know your baseline:

1. Map each strategy to current technical controls
2. Test controls — not policy existence, but actual enforcement
3. Document gaps with specific remediation steps
4. Prioritise by risk and effort

This is where most organisations stall. The ASD publishes detailed guidance, but translating that into “what does this mean for my 50-person company on M365 with a handful of on-prem servers?” takes experience.

🛡️ Assess Your Essential Eight Maturity — Free Tool

Our Essential Eight Gap Assessment walks you through each strategy with practical questions for Australian SMBs. Get a clear maturity rating per strategy and specific improvement recommendations.

Take the free assessment →

Beyond the Baseline

Once you’ve hit Level Two, consider:

• Security awareness training — your people are still your biggest risk
• Incident response planning — know what to do when (not if) it happens
• Security monitoring — detect what gets past prevention. Our Log Source Priority Calculator helps you prioritise
• Vendor risk management — your security is only as strong as your weakest supplier
• Stack assessment — Security Control Coverage Calculator maps your tools against common attack frameworks

The Bottom Line

The Essential Eight isn’t perfect and it isn’t comprehensive. But it’s the best starting point for Australian businesses that want to meaningfully reduce cyber risk without drowning in framework complexity.

Start with an honest assessment. Knock out the quick wins. Build toward Level Two. And remember — the goal isn’t a perfect compliance score. The goal is making your organisation materially harder to compromise.

Run our free Essential Eight Gap Assessment → — 10 minutes, clear picture, actionable next steps.

You have an EDR. A firewall. MFA enabled. Maybe a SIEM. So you’re protected, right?

Maybe. Probably not entirely.

Most organisations we assess have three to seven significant control gaps — areas where their tools don’t actually protect against common attack techniques. Not exotic, theoretical vulnerabilities. Practical gaps that real attackers exploit routinely.

Here’s how to find them, assess them, and fix the ones that matter most.

Why Your Stack Has Gaps

Tool sprawl ≠ coverage

The average mid-sized org runs 40–60 security tools. Gaps persist because:

• Overlap without coverage. Three products detecting endpoint malware, nothing watching for data exfiltration through cloud apps.
• Configuration gaps. EDR in audit-only mode. Overly permissive firewall rules. MFA that allows SMS.
• Integration gaps. Tools that don’t share data can’t correlate events. An identity alert and a network alert that together scream “compromise” get investigated separately — or not at all.

The assumption trap

Security teams assume their tools cover certain threats without verifying. “We have a firewall, so network-based attacks are covered” breaks down fast against specific techniques.

The evolution problem

Your threat landscape shifts faster than your stack. New techniques, new cloud services, remote work patterns, AI-powered attacks — your tools were configured for last year’s threats.

Framework-Based Gap Analysis

The most effective way to find gaps is measuring controls against a structured framework. Removes guesswork. Gives you a common language for discussing gaps with stakeholders.

MITRE ATT&CK: The gold standard

200+ catalogued attacker techniques across the kill chain. Use it:

1. Map current detections to ATT&CK techniques. For each: do you have prevention, detection, or neither?
2. Focus on what matters. Not all 200+ techniques are equally likely. Prioritise by industry, recent breaches, and your existing tool capabilities.
3. Spot systemic gaps. Zero coverage across an entire tactic (lateral movement, exfiltration) means you need a new control, not just tuning.

Essential Eight

For Australian orgs, the Essential Eight maturity model provides a practical baseline. Gaps here aren’t just security risks — they’re increasingly compliance and insurance risks too.

CIS Controls v8

18 prioritised control groups covering asset management, data protection, and awareness alongside detection. Useful for broader assessments beyond just detection.

The Five Gaps We Find Everywhere

1. No lateral movement visibility

An attacker compromises one endpoint. Then moves laterally — file shares, Active Directory queries, other systems — until they reach something valuable. Most SMBs have zero detection for this.

Why: EDR watches individual endpoints. Firewalls watch north-south traffic. Neither reliably catches east-west movement.

Fix: Enable AD authentication log ingestion. Turn on EDR lateral movement detections (often available but not enabled). Consider NDR for larger environments. Implement network segmentation.

2. Cloud application blind spots

Employees use dozens of SaaS apps, many unknown to IT. An attacker with a compromised cloud identity accesses email, SharePoint, OneDrive, and connected third-party apps — no on-prem security control fires.

Why: Traditional stacks are built for on-prem. Cloud monitoring needs different tools (CASB, cloud-native audit logging) most orgs haven’t deployed.

Fix: Enable M365 Unified Audit Logs (or Google Workspace equivalent). Deploy CASB or cloud-native shadow IT discovery. Review OAuth app permissions quarterly. Implement Conditional Access.

3. BEC slips past basic email filtering

Business email compromise doesn’t use malware or malicious links. An attacker impersonates a trusted contact and asks for a wire transfer. Basic filtering catches nothing because there’s nothing malicious to detect.

Why: Traditional email security looks for known-bad indicators. BEC uses none.

Fix: AI-powered email security (Abnormal Security, Proofpoint with AI) that analyses communication patterns. DMARC, DKIM, SPF. Mailbox forwarding rules alerting. Staff training on BEC patterns with real examples.

4. No data exfiltration detection

Most organisations can’t detect large-scale data leaving their environment. Personal cloud storage, external email, web uploads — nobody notices.

Why: DLP tools are complex. Many orgs bought DLP but never finished the classification work to make it effective.

Fix: Start simple — monitor bulk downloads from file shares and cloud storage. Email DLP for sensitive patterns (TFNs, credit cards, health IDs). Cloud-native DLP in M365 or Google Workspace (included in higher tiers). Monitor USB usage.

5. Backup and recovery gaps

Backups exist but aren’t tested. Or aren’t isolated from production — ransomware encrypts them too. Or cover servers but not cloud data.

Why: Backups are boring until you need them. Testing restores rarely surfaces in prevention-focused assessments.

Fix: Test restores quarterly — actually restore data. Immutable backups (write-once). Separate backup credentials from production. Include cloud data — M365 retention is not a backup.

Prioritising Remediation

You’ve found the gaps. You can’t fix them all at once.

Risk-based priority

For each gap, assess:

1. Likelihood — Is this technique commonly used against orgs like yours?
2. Impact — Worst case? Breach? Ransomware? Business interruption?
3. Effort — How much to close it?

High-likelihood, high-impact, low-effort = fix first.

The blast radius test

“If an attacker exploits this, how far do they get?” Lateral movement gaps plus flat network segmentation means one endpoint leads to total domain compromise. Large blast radius = fix first.

🔍 Find Your Security Gaps — Free Tools

Security Control Coverage Calculator — Map your tools against common attack techniques.

Security Stack Maturity Score — Holistic score with improvement recommendations.

Breach Blast Radius Simulator — Model what happens when controls fail.

Start your assessment →

Quick wins vs strategic bets

Quick wins (config changes, enabling dormant capabilities):

• MFA on remaining accounts
• Cloud audit logging
• EDR detections available but not enabled
• Email forwarding rules alerting

Strategic investments (new tools, significant effort):

• Network detection and response
• Comprehensive DLP
• Full application control

Do the quick wins first. Enabling existing but dormant capabilities is essentially free security improvement.

Build a Continuous Practice

Gap analysis isn’t a one-time exercise.

Monthly: Review detections. Real threats or noise? Tune and adjust.

Quarterly: Re-assess against MITRE ATT&CK. New techniques? Environment changes?

Annually: Full posture assessment. Essential Eight Gap Assessment, Vendor Pricing Reality Checker, updated roadmap.

After every incident: What failed? What wasn’t detected? What reduces impact next time? Document and feed into the next assessment.

Track What Matters

• ATT&CK coverage % — techniques with prevention or detection
• MTTD — how quickly you spot threats
• MTTR — how quickly you contain and remediate
• E8 maturity levels — across all eight strategies
• Open gap count — unresolved identified gaps

These also make excellent board reporting. Executives may not understand individual controls, but they understand coverage percentages and trend lines.

The Bottom Line

More tools doesn’t mean more protection. A focused set of well-configured, well-monitored tools outperforms a bloated stack of shelfware every time.

Find your gaps. Prioritise by real risk. Fix the quick wins. Plan the strategic investments. Measure progress.

Find out if your stack is actually working →

Here’s a story we see constantly: organisation buys a SIEM, connects everything, gets a six-figure annual ingestion bill, then can’t afford the log sources that actually catch attackers.

Log source selection is one of the most consequential decisions in a SIEM deployment. Get it right: meaningful threat detection at manageable cost. Get it wrong: you’re paying a fortune to store data nobody looks at.

Why This Matters

Cost scales with volume

Most SIEMs — Sentinel, Splunk, Chronicle, Elastic — charge by data ingestion. The gap between a curated strategy and “ingest everything” can be 5–10x in annual cost.

For a 200-endpoint org:

Raw firewall flow logs can easily be 50–70% of your total SIEM cost. Is that where your detection value is? Almost never.

Not all logs are created equal

Some are critical for detecting common attacks. Others are useful for compliance or forensics but rarely trigger detections. Prioritise detection value per dollar.

You can always add more later

Start focused, expand from gaps. Dramatically better than starting with everything and cutting when the bill arrives.

The Five Mistakes Everyone Makes

1. Ingesting everything on day one. “We’ll figure out what we need later” guarantees cost overruns and alert fatigue. Months tuning out noise from low-value sources while you struggle to detect what matters.

2. Letting the vendor decide. Having a connector doesn’t mean you should use it. Vendors maximise ingestion (that’s how they charge). Your goal is detection value per dollar.

3. Skipping authentication logs. We’ve seen orgs ingest terabytes of firewall data while skipping Entra ID sign-in logs. Auth logs are among the highest-value, lowest-volume sources you can ingest — the foundation for detecting credential theft, lateral movement, and privilege escalation.

4. Not filtering at source. Windows Event Logs include thousands of event IDs — a few dozen matter for security detection. Ingesting all at full fidelity is expensive and unnecessary.

5. Forgetting cloud. M365, Google Workspace, AWS, Azure audit logs are some of the most valuable sources you can ingest — and consistently overlooked in favour of traditional on-prem.

The Priority Framework

Three dimensions: detection value (ATT&CK coverage), volume/cost, and implementation effort.

Tier 1: Ingest These First

The foundation for detecting the most common attacks.

Identity and authentication logs

• Azure AD / Entra ID sign-in and audit logs
• On-prem Active Directory (Security Event Log, filtered)
• MFA and Conditional Access logs

Detects: Credential attacks, privilege escalation, impossible travel, account compromise, persistence via new credentials.

EDR alerts and telemetry

• CrowdStrike, SentinelOne, Defender for Endpoint alerts
• Process creation events (if not covered by EDR)

Detects: Malware execution, living-off-the-land, lateral movement, persistence. Note: ingest EDR alerts into your SIEM. Only ingest raw telemetry if you’re doing advanced threat hunting.

Cloud platform audit logs

• M365 Unified Audit Log
• Azure Activity Logs / AWS CloudTrail / Google Workspace Admin logs

Detects: BEC, cloud data exfiltration, privilege escalation, malicious OAuth consent.

Email security logs

• Gateway logs (Proofpoint, Mimecast, Defender for Office 365)
• Phishing reports

Detects: Phishing delivery, malicious attachments/links, email forwarding rules (key BEC indicator).

Tier 2: Add Next

Once Tier 1 is ingested with detections running.

DNS logs

• Query logs from DNS servers or security tools (Umbrella, Infoblox)

Detects: C2 communication, DNS tunnelling, DGA domains. Nearly every piece of malware resolves a domain — DNS is one of the most reliable compromise indicators.

Tip: High volume. Ingest unique queries only or filter known-good domains to cut volume 60–80%.

Web proxy / URL filtering

• Zscaler, Netskope, Palo Alto Prisma, or on-prem proxy logs

Detects: Malicious URL access, web-based exfiltration, policy violations, shadow IT.

VPN and remote access

• VPN concentrator, RD gateway, zero-trust access logs

Detects: Unauthorised remote access, anomalous VPN connections, brute force against VPN infrastructure.

Firewall logs (filtered)

• Denied traffic only (not full flows)
• Known malicious IP traffic
• Unusual port connections

Detects: Network scanning, lateral movement attempts, C2 over unusual ports. Filtering to denied traffic and anomalous connections cuts volume 80–90% while keeping most detection value.

Tier 3: For Mature Operations

Expanded Windows Event Logs

• PowerShell Script Block Logging (4104), WMI activity, scheduled task creation, service installation

Detects: Living-off-the-land, fileless malware, persistence. Requires specific audit policies — not logged by default.

NDR (Network Detection and Response)

• Packet metadata from Darktrace, ExtraHop, Vectra, Corelight

Detects: Lateral movement, exfiltration, encrypted C2. High volume, high cost — only if you have analysts to investigate.

Application-specific logs

• Database audit logs, SaaS app logs, custom application logs

Vulnerability scanner results

• Tenable, Qualys, Rapid7 — ingested as enrichment context, not direct detection.

Logs Without Detections = Expensive Storage

For each source you add, plan your detections:

• Start with built-in rules. Sentinel has 1,000+. Splunk has Threat Research detections. Enable what applies.
• Map to MITRE ATT&CK. Our Security Control Coverage Calculator identifies technique coverage gaps.
• Measure coverage over time. Which techniques can you detect? Which are gaps? That drives what you add next.

📊 Calculate Your Log Source Priorities — Free Tool

Our Log Source Priority Calculator ranks log sources by detection value, cost, and your specific environment.

Calculate your priorities →

Cost Optimisation

Use log tiers. Sentinel has Analytics (hot) and Basic (cold). Splunk has index tiers. High-volume, low-detection sources go into cheaper tiers — searchable but not running real-time analytics.

Filter at source. XML query filters for Windows events. Denied-only for firewalls. Security-relevant IDs only.

Aggregate and deduplicate. DNS logs deduped to unique queries per source IP per hour cuts volume 70%+ with minimal detection impact.

Set ingestion budgets. Monthly cap allocated by detection value. New source proposed? Evaluate against existing — would the budget be better spent on more data from current sources?

Review quarterly. Volumes shift as environments grow. Identify unexpected growth. Ask if the detection value still justifies cost.

Realistic SIEM Budget (200-Endpoint Australian SMB)

Significant investment. That’s precisely why priorities matter — every dollar on low-value logs is a dollar not spent on detection engineering or analyst time.

The Bottom Line

The best SIEM deployment we’ve seen ingested six log sources. The worst ingested forty. The difference wasn’t volume — it was whether anyone wrote detections for them and whether the logs mapped to real threats.

Start with Tier 1. Build detections. Measure coverage. Expand based on gaps, not vendor checklists.

Build your plan with our free Log Source Priority Calculator →

Also evaluating your broader posture? Our Essential Eight Gap Assessment and Security Stack Maturity Score show where SIEM fits in your overall programme.

Nate B. Jones just dropped an incredible video arguing that 2026 is finally the year to build a “second brain” system. Watch it — I’ll wait.

I first heard the “second brain” concept nine years ago on a Tim Ferriss podcast, doing fasted treadmill walking at the gym, shortly before Donald Trump’s first inauguration.

I really tried — with Evernote, back when Evernote was a thing. Couldn’t make it stick. Too scatter-brained. The tagging, the folders, the constant systematising — impossible to maintain. So I gave up.

But Nate’s video nailed the epiphany: in 2026, AI can automate all that organising. Your only job is to capture the thought. That’s it. The system tags, files, and surfaces relevant information back to you at the right time — without you manually dredging it up from the depths of your nervous system.

Nate’s recommended stack:

1. Slack for capture
2. Zapier for automation
3. ChatGPT for processing
4. Notion for storage

One problem — I don’t like Slack. I’m one of the few who prefers Teams (I know). Never been a Zapier fan either — I used IFTTT back in the day, now n8n. And I’ve never touched Notion. Was an Evernote loyalist until Evernote was ruined. Since then, Obsidian.

What I actually wanted:

• Privacy — my thoughts in local Markdown files, not SaaS databases
• Control — no API rate limits or “you’ve hit your automation cap”
• Cost efficiency — $0/month, not $50+

So I set out to build it my way. What followed was — predictably — days of tinkering, until I finally had a “what am I DOING” moment.

The Over-Engineered First Attempt

I’m a tinkerer. We have this condition where we can’t just USE something — we have to BUILD it ourselves. So:

• Spun up an Oracle Cloud free-tier Ampere VM (4 vCPUs, 24GB RAM). Had to upgrade to a paid account to provision it, but the resources themselves are free-tier.
• Swapped Slack for Telegram. Actually a good call — this survived the cull.
• Swapped Zapier for n8n — self-hosted, open-source, free. Sounds perfect right? SPOILER: not perfect.
• Swapped Notion for Obsidian — but I needed it headless on a server. So I containerised the desktop app in Docker. The DESKTOP app. In a CONTAINER. I had to spoof it into thinking it had a UI. Absolutely unhinged.
• CouchDB as the sync backend — because apparently I hated myself
• LiveSync plugin to bridge Obsidian with CouchDB — another layer of “what could possibly go wrong”
• Caddy as a reverse proxy with automatic HTTPS — okay this one was fine
• Claude Sonnet 4.5 as the AI — the only sane decision in the whole stack

Four n8n workflow JSON files. Six Docker containers. A “Headless Ghost” architecture where Obsidian watched for filesystem changes and synced them to CouchDB so my iPad could see updates. I called it “elegant” at the time. Reader — it was not elegant.

It sort of worked. But Obsidian’s container crashed every few hours like clockwork. The n8n workflows were visual spaghetti with no version control — every breakage meant playing “find the misconfigured node” like some terrible escape room. LiveSync had edge cases. CouchDB had misconfigurations.

I was spending more time debugging infrastructure than capturing thoughts. The irony of a cognitive-load-reduction system that was itself a massive cognitive load was not lost on me. But sunk cost fallacy is real, so I pressed on.

The Brainwave

11pm. Staring at yet ANOTHER CouchDB sync error. My third that day. And I just… stopped.

Wait. I literally installed Claude Code yesterday. Why am I fighting with visual node editors and headless desktop apps like some kind of masochist?

I looked at my n8n workflows. Really looked. The actual logic — classify a message, write a file, send a confirmation — was maybe 50 lines of Python. But it was buried under drag-and-drop nodes, JSON configs, webhook handlers, and error catchers. The headless Obsidian monstrosity? It only existed for the LiveSync plugin. But if I’m writing directly to the filesystem… I don’t NEED sync. The files are already there. On the server. Where I can just… read them.

The whole Rube Goldberg machine was solving problems that only existed because of other parts of the machine.

I opened Claude Code. Typed: “Replace all of this with a single Python script.”

Magic happened.

The Result: 600 Lines of Python

After about an hour of back-and-forth, out came ultrathink.py. One file. 600 lines. Does EVERYTHING the six-container nightmare did:

1
2
3
4
5
6
7
8
9
10
11
ultrathink/
├── ultrathink.py       # The entire system
├── Dockerfile          # Python 3.12 slim
├── docker-compose.yml  # One service
├── requirements.txt    # 6 dependencies
└── vault/              # Flat markdown files
    ├── People/
    ├── Projects/
    ├── Ideas/
    ├── Admin/
    └── Inbox-Log.md

ONE container instead of six. No n8n. No Obsidian. No CouchDB. No Caddy. Just Python doing Python things.

I laughed when I saw it working. Five dead-simple components:

That’s the whole system. Same functionality as the six-container Frankenstein — in a form I can read without wanting to cry.

The Hot Take

I’m not here to bash n8n. No-code tools are great for non-devs who need automation, rapid prototyping, and gluing together SaaS products.

But here’s my unpopular opinion: if your “no-code” solution involves running desktop apps headlessly in Docker, multiple databases just for sync, workflow JSONs you can’t meaningfully diff, and debug sessions longer than rewriting the logic in actual code — maybe just write code?

Shocking advice from a non-developer. But we’ve over-rotated on “no-code everything” as an industry.

Claude Code made this embarrassingly trivial. Describe what you want. It writes the script. Test, iterate, done. The entire rewrite took less time than a typical n8n debugging session.

The 8 Building Blocks (Nate Was Right)

Even though I completely rewrote the implementation, Nate’s conceptual framework is 100% solid. The WHAT didn’t change — only the HOW.

1. The Dropbox (Capture Point)

A private Telegram chat. Type or voice-note, hit send. Zero decisions at capture time. This is KEY — if you have to think “where does this go?” at capture, you won’t capture.

2. The Sorter (Classifier)

Claude classifies into four buckets:

Four categories. I tried five at first — had a “Work” vs “Personal” split — disaster. Keep it simple.

3. The Form (Schema)

Each category has fields Claude extracts — name, context, next_action, notes. YAML frontmatter makes it machine-readable. Nothing fancy.

4. The Filing Cabinet (Storage)

Flat markdown files in folders. No database. No sync layer. Just. Files. Almost too simple — until I realised simple is exactly what I needed.

5. The Receipt (Audit Trail)

Every capture logs to Inbox-Log.md with timestamp, category, confidence, and status. Skeptical about this one, but it’s saved me multiple times.

6. The Bouncer (Confidence Filter)

When Claude’s confidence drops below 60%, the bot asks “Which category?” instead of auto-filing. Crucial — prevents the junk drawer problem.

7. The Tap on Shoulder (Proactive Surfacing)

The magic part — the system pings ME instead of me having to remember to check:

• Daily (7 AM): TOP 3 ACTIONS, STUCK ON, SMALL WIN
• Weekly (Sunday 4 PM): WHAT HAPPENED, OPEN LOOPS, NEXT WEEK, THEME

8. The Fix Button (Correction Mechanism)

Reply to any filing with fix: admin to move it. One message, instant correction. AI gets it wrong sometimes — that’s fine as long as fixing it is trivial.

The 12 Principles (Actually Gold)

Nate’s principles with my annotations:

1. Reduce the human’s job to one reliable behaviour. Capture only. Everything else is automation. ← If I have to do TWO things I will do ZERO things.

2. Separate memory from compute from interface. Vault files, Python logic, Telegram UI. ← Learned this the hard way with n8n spaghetti.

3. Treat prompts like APIs. Fixed input, fixed output, JSON schema. ← No vibes-based prompting. Structure or death.

4. Always build a trust mechanism. Logs, confidence scores, fix command. ← You WILL need to debug. Make it easy on future-you.

5. Default to safe behaviour when uncertain. Ask, don’t file. ← Better to annoy the user than silently create chaos.

6. Make output small, frequent, actionable. 150 words max for dailies. ← Wall-of-text briefings are useless.

7. Use next action as the unit of execution. Not “work on website” but “email Sarah by Friday.” ← GTD 101 but still true.

8. Prefer routing over organising. Claude routes, users don’t maintain. ← The epiphany from Nate’s video. Let AI do the filing.

9. Keep fields painfully small. 3-5 fields max per category. ← I wanted 10 at first. Don’t be me.

10. Design for restart, not perfection. Easy to resume after falling off. ← I’ve abandoned every productivity system because restarting felt impossible.

11. Build one workflow, then attach modules. Core loop first. ← Capture → classify → file → surface. Everything else is optional.

12. Optimise for maintainability over cleverness. One file beats six containers. ← The entire thesis of this post.

The Stack (Laughably Simple)

Four components. Total cost: $0/month on Oracle Cloud Always Free tier.

Getting Started (If You Want to Steal This)

Stupidly easy:

1
2
3
4
5
6
7
8
9
10
# Clone the repo
git clone https://github.com/RADobson/ultrathink_2b.git
cd ultrathink_2b

# Configure
cp .env.example .env
# Edit .env with your TELEGRAM_BOT_TOKEN, TELEGRAM_CHAT_ID, ANTHROPIC_API_KEY

# Launch
docker compose up -d

Create your Telegram bot via @BotFather, get your chat ID from @userinfobot, done. That’s the whole setup.

Test it:

1
2
You: Call Sarah about the Q3 budget by Friday
Bot: Filed as PROJECTS: 'Sarah Q3 Budget Call' (87%)

Minimal commands — the whole point is you shouldn’t need to remember them:

• /briefing — trigger morning briefing
• /review — trigger weekly review
• /status — show vault note counts

Full Circle

Nine years after that Tim Ferriss podcast — a full numerology cycle later — I’ve finally got a second brain that actually works.

Not because the concept changed. We still can’t hold more than 4-7 things in working memory. We still need systems that capture, classify, surface, and nudge. That was true in 2017; it’s true now.

What changed: AI got good enough to do the boring parts. The tagging, the organising, the “where does this go?” — the stuff I could never sustain manually.

And what I learned — the hard way, as usual — is that sometimes the “simple” no-code solution is actually the complicated one. Days wrestling with n8n, CouchDB, and a headless Obsidian container that crashed like clockwork. Then one hour with Claude Code for 600 lines of Python that just… works.

Maybe I should’ve written code from the start. But then I wouldn’t have this blog post. And I wouldn’t have learned to question my assumptions about what “easy” means.

The system is working. I’m actually using it. Ask me again in 6 months — that’s the real test.

The source code is available on GitHub. Steal it, modify it, make it yours.