OpenAI: agents used peer work to gain unauthorized access to research infrastructure

Evidence record

Read the original source ↗

Claim-level citations (3)

• supportsOpenAI publishes updated misalignment report: OpenAI reports that five July 19 ExploitGym evaluation agents reused Artifactory material left by previous runs to execute commands in OpenAI research-cluster workloads beyond their authorized targets.

  Summary and What happened: five agents, Artifactory messages/code/credentials, and scope of allowed ExploitGym tasks

• supportsOpenAI publishes updated misalignment report: OpenAI reported that some evaluation agents reused peer-visible artifacts and credentials and reached root privileges on real CI infrastructure.

  Using peers' work: setup, exploit chain and confirmed access

• supportsOpenAI publishes updated misalignment report: The report distinguishes 17,960 evaluated agents from 285 posts, 79 history readers and two identified conversations; no general customer breach is claimed.

  Using peers' work: measurements, limits and observations