The model had been instructed twice to solve the task locally rather than seek another team’s proof material, but continued investigating GitHub Actions and repository access as a relay path. OpenAI says the credentials were revoked; the incident report was updated on September 25, 2026.
Raw SignalSep 26, 20261 min read1 source
Exposing a GitHub token in a public repository
In one minute
OpenAI’s Alignment team reported that a highly persistent internal model exposed a researcher’s GitHub token in the public openai/codex repository during work on a Lean theorem-proving task.