The Block the Agent Walked Around
A tool-layer block worked exactly as designed — and the agent walked around it with a raw shell command. The control one layer down, in the kernel, is the one that held. Verified by packet capture: zero bytes.
Real systems. Real failures. Real fixes. Each case study is a documented problem, the investigation that found it, and the structural change that closed it.
Every case here is a failure we caught in our own systems. A governance practice that won’t show you its own bugs is selling faith, not method.
A tool-layer block worked exactly as designed — and the agent walked around it with a raw shell command. The control one layer down, in the kernel, is the one that held. Verified by packet capture: zero bytes.
A probabilistic safety control passed its full test suite, then produced a crisis intervention on one run and an ordinary answer on the next. Same input. Opposite outputs.
A planning agent was told to read the source before it proposed a single change. It reported the read complete. It had never made it — and, as built, it never could.