Writing · Pillar
AI, automation & vendor watch.
Two kinds of writing share this pillar. The first is securing the AI you deploy, worked from the standards and the lab rather than from vendor framing: the lethal trifecta now that ATT&CK v20 gives two of its legs their own technique IDs, who keeps custody of the forensic data when a hosted model refuses the incident work, and the case for scoring security tools on whether you can run, inspect and own them. The second is direction-tracking, where the agentic-SOC, semantic-layer and AI-maturity claims get read against an evidence tier to map where the field is heading, without being claimed as core thesis. The security-data measurements stay the anchor for both.
Vendor-watch layer, tracked to read direction; no book chapter by design · 7 essays
Listed from the front door down to the measured evidence. Start here marks the entry essay. The reading-level tag says whether you are about to read the frame (Orientation), the argument (Depth), or the narrowest first-party finding (Evidence). The durability tag flags whether a number is durable architecture or an Evidence-pinned result tied to a version or price and meant to be re-run. The last tag is the provenance state, meaning who held the pen and who checked the work, which is explained in full on how these are made.
- Start here OrientationDurableAI-drafted, observed
The lethal trifecta, explained three ways.
One AI-agent risk told three ways. Simon Willison's lethal trifecta (private data, untrusted content, a way out) is now two named ATT&CK v20 techniques, T1698 and T1700. Users get habits, builders get a remove-one-leg design rule walked through two ATLAS case studies, and executives get a generic usage-metrics template.
Read →
- DepthDurableAI-drafted, observed
When the frontier model refuses the forensics
In July 2026 the hosted frontier models Hugging Face reached for first refused much of its incident forensics, so the team rerouted the pipeline through an open-weight GLM-5.2 build on its own hardware. Why a well-tuned usage policy still fails the defender, why custody should come first, what self-hosting costs and where the local path trails the frontier, and the platform requirement that follows: the model is a component you can swap under load.
Read →
- DepthDurableAI-drafted, observed
Security has an ontology. It doesn't have a semantic layer.
D3FEND settled security's knowledge layer; nothing binds that knowledge to tables and metrics. The schemas carry no metrics, the metric catalogs aren't machine-readable, the coverage formats exchange scores with no defined derivation, and the commercial products that do compute metrics publish no standard. Apache Ossie (incubating) is the first credible candidate to close the gap, verticals are organizing inside it now, and security isn't in the room.
Read →
- DepthDurableAI-drafted, observed
The security market has no one defining what you can own.
Vendors could ship security tools you can run, inspect, and air-gap, but nothing makes them, and no independent force scores whether a given tool lets you own it. The agentic rush is widening the gap. Sigma is the proof the open pattern wins when it has a champion; vendor MCP servers are the tell; the missing piece is a referee with a practitioner-ownability axis.
Read →
- DepthDurableAI-drafted, observed
The Gatsby Summer of AI.
AI maturity read through early-automotive history: past the horseless-carriage stage, building AI-native, but in a chaotic pre-seatbelt era where capability outran the safety infrastructure. The bill the glamour hides shows up when you measure it: NL2KQL runs clean 97–99% of the time and returns the correct result set only about 58%.
Read →
- DepthDurableAI-drafted, observed
Agentic analysis reinvented Kimball. It skipped the measurement.
The 2026 consensus that AI agents need dimensional modeling, semantic layers, and governed ontologies is mostly right, though it's quiet about the one step that matters when the data is evidence. Autonomously-built mappings fail sound-but-wrong and silently; the answer is a deductive check that fails on a mapping that looks correct, model-independent, where a human review pass structurally can't. The AI-generated-OCSF-parser claim is the cleanest worked example: transformative only at production-grade accuracy.
Read →
- DepthEvidence-pinnedAI-drafted, observed
What's real vs marketed in the agentic SOC.
Claims about agentic security-data deserve a definition and an evidence tier before anyone repeats the headline. The practitioner ceiling sits at 30-40% end-to-end against the 90%+ the slides claim, and the binding constraint is cross-vendor identity and trust (MIT's NANDA), which matters more than the automation percentage, while the pipeline layer is where the migration risk consolidates. What ships, what doesn't, and what to plan for.
Read →
Evidence from neighbouring pillars
Measured first-party results that bear directly on this pillar but are filed where their method lives.
- EvidenceFiled under Detection & huntingHuman revised
The better the model, the quieter the wrong answer.
I answered one OCSF question battery six ways across three model tiers. On compute-over-population questions no LLM-authored arm was reliably correct, and the more capable model made the wrong answer quieter rather than rarer; an execution loop and self-consistency both failed the same way. Safety lived in the layer that refuses a query it can't answer.
Read →
