Reproduce “Screwtape: Adversarial Transcript Editing for Monitor Red-Teaming” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Gosha Kozhevnikov, Tyler Tracy
Screwtape edits real coding-agent transcripts to test action-only monitors. An attacker revises edits using feedback from a verifier checking plausibility and severity, and a proxy monitor. Across 252 runs on 94 Redwood staff transcripts, human review confirmed 6 of 17 attacks missed by a held-out developer-provided monitor; the edits are unexecuted stress tests, not deployment forecasts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.