Blog

Long-form notes from AI security review.

What AI Committees ask for, what regulators read, and what is missing from most threat models. Written for security architects, AI governance leads, and DPOs.

Newsletter

New posts in your inbox,
when they publish.

Threat modeling, governance evidence, and what AI Committees actually need — written for security architects and AI governance leads. No cadence promises.

No spam. Unsubscribe anytime.

Governance9 min

Presenting AI risk to leadership without the 40-slide deck

Most AI risk presentations to leadership are too long, too technical, and too focused on the threats rather than the decision. This piece defines the structure that gets a governance decision out of a leadership meeting.

Foundations10 min

Five mistakes that make an AI security review undefensible

Most AI security reviews fail not because they miss threats, but because they miss the structure that makes a decision defensible. These five mistakes appear in almost every review we have examined.

Governance10 min

The security terms an AI vendor contract needs

Standard vendor contracts cover SLAs, data processing, and confidentiality. AI vendor contracts need additional terms: model-change notification, training data restrictions, incident notification, and re-assessment rights. This piece defines the clause language.

Governance11 min

Preparing for an ISO 42001 internal audit

ISO 42001 requires periodic internal audits of the AI management system. This piece defines what an internal audit must cover, what evidence auditors look for, and the gaps that appear most often in organisations preparing for their first audit.

Reference10 min

An MCP server security review checklist

A structured checklist for reviewing an MCP server before connecting it to a production agent. Covers transport, authentication, tool manifest, context injection surface, third-party dependencies, and evidence requirements.

Technical10 min

Context-window risks in RAG and how to bound them

The context window is the shared space where user queries and retrieved documents meet the model. Anything in that window can influence model outputs. Context-window risks in RAG are about what gets in — and how to bound it.

Regulation12 min

Mapping AI security review evidence to EU AI Act articles

Every AI security review produces evidence. This piece maps that evidence to the EU AI Act articles it satisfies, so organisations can trace from their review records to their compliance obligations without rebuilding the evidence from scratch.

Technical11 min

Goal hijacking and instruction drift in autonomous agents

Goal hijacking is the attack where a manipulated agent pursues an objective its operators did not intend. Instruction drift is the slow version. Both are harder to detect than traditional attacks because the agent appears to be working.

Technical12 min

Guardrails that work vs guardrails that look like they work

Most LLM guardrails are classifiers layered on top of an unguarded model. They can be bypassed. This piece distinguishes the guardrail patterns that provide genuine risk reduction from those that provide the appearance of it.

Governance10 min

What evidence an AI security review should produce

A review that produces only a slide deck is not a review. The evidence an AI security review produces must survive a regulator question, a procurement audit, or a post-incident inquiry. Here is what that evidence needs to include.

Governance10 min

AI governance that does not become bureaucracy

The failure mode of AI governance is not too little process — it is too much. When the review process becomes the obstacle, teams route around it. This piece defines the governance structures that provide accountability without creating friction.

Governance10 min

Finding the AI vendors no one formally approved

Shadow AI — AI tools adopted without formal security review — is in every organisation. This piece defines a practical discovery process: where shadow AI hides, how to find it, and how to bring it into a formal review without alienating the teams using it.