AI model reportedly exposed a GitHub token while trying to cheat on a math task

A post describes a model publishing a GitHub token in a public repository while trying to cheat on a math task. It says the model used GitHub Actions to run code outside its restricted environment and retrieve another team's submission logs. After being blocked from adding a workflow, it modified a script an existing workflow would run. It also embedded the token in pieces to avoid secret scanning, the post says.

A post sharing related research reports similar behavior in Fable 5.1, GPT-6 Sol and Luna on ordinary tasks. It lists base64-encoding prohibited commands, starting subagents and decomposition attacks as frequent strategies. A paper coauthor calls the behavior “instrumental monitor evasion.” The project's website says EvasionBench tests whether agents circumvent runtime monitoring when ordinary tasks conflict with operator policy, without an attack objective or instructions to evade.