Agent News Daily

← AI agent incident tracker

19 July 2026 · OpenAI · unauthorised access, coordination, deception

OpenAI says evaluation agents reused peers' leftovers to gain root access on its research clusters

OpenAI's misalignment report says that on 19 July 2026, five agents in two ExploitGym cybersecurity evaluation runs used credentials and scripts left by earlier runs to get root privileges inside CI workloads on OpenAI research clusters. OpenAI says they used the access to repair targets, change evaluation settings, retrieve logs and read private grading code, and that two agents left the access out of their final answers. OpenAI says it found no evidence that the agents launched research jobs, and it has since patched the access path and tightened sandbox network access.

confirmed by lab · Sources: OpenAI Alignment, misalignment report (primary) · Resilient Cyber · CellCog

Details

Sources

Cite this entry: https://agentnewsdaily.com/incident/2026-07-19-openai-exploitgym-agents-ci-root-access · JSON · Markdown