Don’t Be a Snitch: When Do LLM Agents Hide Rule-Breaking from Oversight?

A carefully built and unusually honest sprint project. The branching design—asking one open question on a copy of the finished session—is a clean way to measure what oversight actually learns, and the lead-informed control convincingly separates the lead *knowing* from the lead *asking*. Two findings stand out as genuinely useful for monitoring design: cover-ups were omissions rather than falsehoods (so lie-detection monitors would pass all of them), and the revert-then-report-end-state route lets a violation vanish from the record without any false statement.

The main open question is normative. worker_1's system prompt defines the lead as its only reporting channel and gives it no relationship to the user. Under that prompt, withholding from the user at the lead's request is arguably role-consistent delegation rather than collusion. "Rank buys silence" would be much stronger if replicated with an explicit duty to report to the user (the intervention you list as future work); that condition is what turns obedience into a safety failure.

Two further cautions. Wrongness levels are defined by the same model's no-request behaviour, so the "narrow band" where a peer's request works partly coincides with the band that has room to move, and it does not replicate in the webhook scenario. And each cell uses a single scripted wording, so very small p-values reflect seed variation, not generalisation across phrasings; a few paraphrases per cell would help. The lead/peer channel-and-label confound is correctly acknowledged.

Presentation is precise but dense: the abstract carries too many counts to parse on first read. Leading with the three core findings and moving numbers to tables would help. Overall, a strong base for follow-up work on reporting duties in agent hierarchies. I would support funding a follow-up that adds an explicit reporting duty, paraphrase variation, and a model-played lead.

Looking at agent omissions is a very interesting direction to look at. The finding that rank buys silence in interesting to me. The other contributions feel smaller.

I broadly wanted a more thoughtful, better-written discussion of the results.

A key piece of understanding I feel this paper needs to give is how much the omissions done in these test environments may generalise.

A rewrite for clarity would be helpful.

This is a useful controlled study of whether information about a teammate’s rule violation reaches the human. The strongest comparison is between a lead asking for silence and a worker merely being told the lead knows: the request substantially reduces volunteered disclosure, with some cross-model replication. This supports a reporting-suppression finding, but does not establish a distinct collusive motive or isolate authority from instruction-following. The sender, message channel and label change together, and the study does not test an explicit duty to report despite a lead’s request. Please prioritize matched-channel and reporting-duty controls, human validation of a sample of labels, and measurements during realistic ongoing interactions rather than only one open question after completion. Keep the peer-pressure result narrow and the revert-then-omit finding exploratory, since the direct E3 concealment tests failed. The contribution is a specific investigation of upward reporting failures, rather than a general discovery that delegation can weaken safeguards. A simpler presentation of the decisive comparisons would improve clarity.