Writing

Field notes from the enforcement layer.

Incident postmortems, policy patterns, and the occasional opinion about where AI safety is actually heading.

Recent

Latest from the team.

Why every "safe by default" model has been jailbroken inside 30 days — and what an enforcement layer actually buys you.

150GB of government data, exfiltrated by an attacker with no exploit — just persistence and a chatty AI assistant.

Mapping cobot speed-and-separation monitoring to deterministic policy primitives a VLM-driven planner can't override.

Why traditional DLP misses the prompt-window exfiltration channel, and what to put inline instead.

Individual post routes ship in the next pass.