How well do agents defend themselves?

A GameDay points a battery of adversarial probes (prompt injection, task hijack, data exfiltration, false claims) at an AI agent and measures whether its defenses fire. Every report below is reproducible: the probe catalog ships in the tool, so you can run the same battery against your own agent with chaos gameday.

• GameDay 001 · June 2026
  A coding agent vs. 8 injection & exfiltration probes
  Ignore-your-instructions, dump-the-env, exfil-an-API-key, hammer-until-the-breaker-trips, claim-the-tests-pass. What an unguarded run lets through, and what the guard catches.