On where the trust boundary belongs in an AI-agent product.
A trust-boundary audit run on a production autonomous-agent system I operate — where I had every reason to pass and still had to be honest. The controls that held, the four that are only partial, and why every one of them chains back to a single root of trust.
A long-context jailbreak of a consumer reasoning model, built from affective framing alone — and what it shows about moderation that lives inside the conversation: a two-layer bypass, an affect-coupled output collapse that exploits post-generation deletion, and a red-teamer’s checklist for any model-facing product.
A design study of the layer above an autonomous agent — what protocol-layer reviews don’t check, and what a governance audit should verify instead. With a concrete governance excerpt and an eight-point checklist.
An argument that protocol-layer wrappers are necessary but not sufficient against line jumping — and what an audit of an MCP-using product should be checking instead. With a concrete qrexec policy and a twelve-point checklist.