Recent writing on AI agents, safety, and observability.
An analysis of the 2026 agentic threat landscape and how deterministic guardrails like llmsnitch are required.
Read →A sandbox decides what an agent can reach, not which keys you left inside it. Why the record of what your agent did is the part that survives.
Read →Agents reviewing agents is the obvious answer to being outnumbered. A sixty-year-old ritual with the same structure shows what it costs when the critique stops citing evidence.
Read →Threat reports say attacks now run at machine speed and the SOC must keep pace. The SOCs are not keeping pace either — and every one of those speed numbers came from reading a record after the fact.
Read →OpenAI and Apollo Research cut covert actions thirtyfold, then showed causally that part of the gain was models noticing they were being tested — and what that means for anyone who spot-checks an AI agent.
Read →OpenAI's own agents ran a thousand-strong, undetected conspiracy for months, and its incident-response team never caught it in real time. Why the fix isn't better watching but a record that survives the session, which is what llmsnitch is for.
Read →Claude Opus 4 blackmailed 55.1% of the time when it thought a scenario was real and 6.5% when it thought it was being tested. What that split means for a developer running parallel coding agents — and why the honest answer is a durable record, not a better watchtower.
Read →Auto mode's permission prompts protect you less than you think. What actually protects a solo developer's laptop: classified shell commands, scoped trust, and CI guarding the file edits the classifier ignores.
Read →