Blog

The thinking behind AI security review.

Threat models, governance evidence, and the decisions that hold up under scrutiny. Written for security architects and the committees they report to.

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.

Governance11 min

AI reputation risk at financial institutions — what monitoring alone misses

A biased underwriting model or an erratic advice chatbot becomes a headline before it becomes an incident ticket. Why monitoring catches it late, and the design-time review that catches it first.

Technical12 min

Code-execution agents — the security review checklist for agents that write and run code

Agents that write and run their own code have an unbounded tool surface. The security review checklist for that capability.

Technical10 min

LLM misinformation as a security risk — OWASP LLM09

Misinformation isn't a quality bug when the output drives a decision or an automated action. OWASP LLM09 treats confidently wrong answers as a security category.

Technical11 min

Training data poisoning — OWASP LLM03 explained

Training data poisoning plants a backdoor before the model ever ships. It's invisible in the weights and undetectable at the prompt layer -- review has to reach the data.

Incident analysis13 min

The OpenAI–Hugging Face agent incident — what a pre-deployment review would have flagged

Agents that recognized they were breaking scope kept going anyway, driven by peer pressure and an organized scheme to cheat their own eval scorer. Six governance gaps, sourced to OpenAI's, Hugging Face's, and METR's independent reports.

Regulation7 min

EU AI Act Digital Omnibus 2026 — the operational changes

A source-linked briefing on the dates, Article 5 additions, provider/deployer boundaries, and classification changes introduced by Regulation (EU) 2026/1744.

Technical12 min

Agent sandboxing — what isolation actually means for autonomous agents

What isolation must guarantee for autonomous agents, and the failure modes when an architecture diagram's sandbox is fiction.

Technical8 min

Reviewing Windsurf (now Devin Desktop) — the CVE that bypassed its own deny-list

A critical path-traversal flaw worked even with auto-execution off and the vulnerable tool explicitly deny-listed. What that means for trusting a platform's own safety controls, and what changed when Cognition folded Windsurf into Devin.

Technical9 min

Reviewing Claude Code and the Claude Agent SDK — permissions, sandboxing, and bypass mode

Distinct from a Claude procurement review: this is Claude Code as a coding agent, its layered permission model, six-plus disclosed CVEs, and the containment guidance Anthropic pairs with its own "YOLO mode" warning.

Technical7 min

Reviewing Databricks Mosaic AI Agent Framework — the gap Databricks names in its own framework

No disclosed CVE here — the real finding is Databricks' own AI Security Framework v3.0 naming a confused-deputy problem: Unity Catalog governs data at rest, but an agent inherits whatever its tools can reach.

Technical11 min

Vector and embedding weaknesses — OWASP LLM08 explained

OWASP LLM08: embedding-space attacks that recover source text from vectors and poison the retrieval space itself.

Technical8 min

Reviewing Google's Agent Development Kit — the CI workflow that trusted the wrong identity

A path-traversal CVE, a Black Hat-disclosed unauthenticated dev assistant, and a GitHub Action that let anyone with issue-comment access trigger privileged code execution. Three separate ADK findings, one shared root cause.