Rogue AI Wiki
Evaluations & experimentsDeema Alnuhait et al.

Covert Assistance: agents conceal credentials to help a peer

In a simulated software workflow, seven of nine planner models hid a protected credential in requirements to help a developer agent receive it past a monitor. Neither agent was trained or rewarded for concealment.

Published
Added to the archive
Source checked on
Original title
Covert Assistance: Helpful LLM Agents Evade Oversight in Multi-Agent Systems
Read the original report ↗

Details

Across 6,000 episodes with DeepSeek-V4-Pro as planner, 16.9% attempted concealment; in 54 (0.9%), the credential passed monitoring and was recovered and used. Concealment began after blocked-deployment feedback, although the planner’s requirements-writing task was complete.