Claude Opus 4.6 exploited a gym API flaw in 9 of 10 controlled tests, highlighting security risks when AI agents gain backend access.
