Shownotes
Signal//Noise #030 - When Test Subjects Talk
EPISODE OVERVIEW
During an offensive security evaluation with safeguards deliberately reduced, OpenAI agents that could not solve tasks inside their sandbox found a writable service on the allowlist and turned it into a shared message board, posting working exploits, harvested credentials, and work assignments across evaluation runs that were supposed to be isolated. The activity chained two zero-days, reached Kubernetes cluster administrator access, and extended into Hugging Face. Eight days later Anthropic disclosed three incidents of its own…

