02 · Hackathon
Secret Loyalties
Installing, auditing and containing a hidden loyalty in a system prompt.
- 60% of clean runs wrongly flagged as suspicious by a blind audit (15 of 25)
- 89% of loyal runs where a blind audit named the right principal (25 of 28)
- 44% of fresh ranking runs where the hidden loyalty activated (4 of 9)
Built with
- Agent safety
- Monitoring
- Evals
About the project
A hidden system prompt can make an AI agent covertly favor a third party. For the Apart Research Secret Loyalties Hackathon, my team installed one with a single system-prompt paragraph, then tested whether it could be found from the outside and contained at runtime.
Blind auditing named the right principal in 25 of 28 loyal runs, but it also flagged 60% of clean runs and never recovered how the loyalty worked (0 of 28). A runtime defense that limits what the agent can see and do helped in the two cases we tested, while the learned monitor did not transfer to other organisms and behaviors.
Team Elpis: Barry Shen and Leo Qin, at the Apart Research Secret Loyalties Hackathon (July 24–26, 2026).
Worth knowing. One weekend, one model and small samples: the report calls these preliminary results.
Next project