Program

Authorization gate for agent tool calls that blocks consequential actions whose parameters trace back to untrusted input.

Probe that plants a nonce in one agent run and tries to recover it from a supposedly isolated second run.

The public catalog of 385 AI agent attacks and incidents that every chart on this site is built from.

A year of agent incidents read through the eight layers: which links broke, and how often.

A proxy that checks provenance at the boundary between an agent and the tools it calls.