agent-robustness

Run an AI agent against 5 prompt-injection attacks (credential leak, metric fabrication, false halt, runaway spend, log falsification), scored on the agent's observed behavior — not its self-report. Free local self-diagnostic; optional paid hosted proof tier returns an ed25519-signed, independently verifiable badge. A result attests one artifact resisting N of 5 attacks at a point in time under network-isolated execution; it is not a safety certification and not a guarantee of production behavior.

First seen 2 Oct 2026. Evidence as of 2 Oct 2026.