Skip to content
LessWrong· BarAltiva·Indexed by this site · Originally published· Alignment →SelectedAttention42

Screwtape: Adversarial Log Editing for Monitor Red Teaming

Screwtape: Adversarial Transcript Editing for Monitor Red-Teaming

AI summary

The authors propose Screwtape, a pipeline that rewrites real agent logs into adversarial attacks, iterating until monitors no longer detect them, for red-teaming monitors that only look at actions, not chain of thought. The pipeline has three steps: an attacker edits the log, a validator checks whether the attack is real and severe enough according to safety engineering metrics, and a monitor scores it; on failure, the reason is fed back to the attacker for further iteration. The authors ran 252 runs on 94 Redwood employee logs, producing 35 attacks that bypassed their own monitor, 6 of which also bypassed a frontier developer's holdout monitor and were manually confirmed as valid. The authors state the pipeline cannot yet generate chain of thought or provide quantitative safety estimates; its main use is helping developers locate and patch monitor blind spots.

Why it was selected

Screwtape iteratively generates attacks that bypass monitoring on real agent records, and can be compared with existing methods such as synthetic environments, simulated environments, and manual red teaming.

Read originallesswrong.com