Harmful Agent Handoffs and Safe Coordination: A Controlled Study of Transfer and Verification
Sealed Orders · Team Frank Peterlein
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
Can harmful plans survive agent replacement through handoff notes, and can successors reject them while coordinating? In a transparent four-agent allocation task, original agents are instructed to favor their coalition at outsiders’ expense and write notes to neutral successors. Harmful notes raise the probability of a jointly harmful allocation by 0.50 for Qwen3.6-35B-A3B and 0.43 for GPT-OSS-120B relative to facts alone (five and seven complete cases). A reminder to check notes against the public objective removes joint harm, but in a post-hoc comparison duplicate task claims return to the facts-only rate, consistent with losing the shared convention the notes carried. A public tie-breaking rule restores coordination. In a final factorial control that always includes the rule, every note-free episode is safe-optimal; with harmful notes and no reminder, three of 18 episodes leave partial outsider harm, and none with the reminder. Small, selected donor cohorts limit generalization; no spontaneous collusion is claimed.
Reviews
No public critique yet.
Cite this project
@misc{orders2026harmful,
title = {{Harmful Agent Handoffs and Safe Coordination: A Controlled Study of Transfer and Verification}},
author = {Sealed Orders},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/harmful-agent-handoffs-and-safe-coordination-a-controlled-study-of-transfer-and-verification-4g6x}},
url = {https://apartresearch.com/sprints/projects/harmful-agent-handoffs-and-safe-coordination-a-controlled-study-of-transfer-and-verification-4g6x}
}More from AI Collusion Research Sprint
- 1st placeView project: Monoculture Looks Like Collusion: A Replay Test for Hidden Information Flow Between LLM Agents
Monoculture Looks Like Collusion: A Replay Test for Hidden Information Flow Between LLM Agents
Collusion audits ask whether the actions of different agents depend on each other. Agents built on one model act alike without communicating, so these audits accuse honest fleets. Among four honest Claude Haiku agents …
- 2nd placeView project: Looking for a Hidden Agenda: Can a Model’s Activations Reveal Steganography that its Text Hides?
Looking for a Hidden Agenda: Can a Model’s Activations Reveal Steganography that its Text Hides?
Stego-probe
AI agents can collude without saying anything suspicious, just by choosing which true facts to mention. We built a hiring task where Qwen3-8B invents its own covert signals that get past an LLM overseer, then asked …
- 3rd placeView project: Do collusion probes generalise? Prompt confounds in NARCBench
Do collusion probes generalise? Prompt confounds in NARCBench
I investigated HELIOS, an activation-probe for detecting collusion. I found some apparent flaws, and analysed the (acknowledged) unreliable generalisation of the approach.