Agent transcript holding a secret

[commjoen]

Context

• What should the challenge scenario be like?
  Coding-agent session transcripts: tools like Claude Code, Cursor, and Aider can leave local session/history files behind. A secret pasted during debugging can end up in one of these files and later get committed or included in a support bundle. Different from a shared-chat leak because the local artifact is the problem.
• What should the participant learn from completing the challenge?
  reveiw transcripts before saving them, AI can make mistakes.
• For what category would the challenge be? (e.g. Docker, K8s, binary)
  AI

Did you encounter this in real life? Could you tell us more about the scenario?

Yes

[kekubhai]

I’d like to work on this one. The local agent transcript angle is interesting and nicely different from the shared-chat leak I’d implement a realistic coding-agent session artifact containing a leaked secret, add the challenge and tests, and document the risk of committing or sharing these transcripts.

[commjoen]

it's yours! Please try to make sure the transcript is "real" (e.g. generate it with codex/cursor/claude so that it actually feels/looks real and not faked)

[kekubhai]

yeah sure , on it