Collusion-Deterring Audit Mechanism for Untrusted Monitoring in AI Control

We design a collusion-deterring audit mechanism for untrusted monitoring in AI control. In untrusted monitoring, a copy of a capable but untrusted model is used to check the outputs of the model itself, as no trusted monitor is as capable. However, the copy might collude: it can behave truthfully on test inputs and wave through an attack that it recognises as its own. The defender holds a small budget of human labour to spend on honeypots, i.e., planted attacks that the monitor has to flag, and on audits of real outputs. We consider that the monitored model is strategic and might collude through its monitor, and propose an audit mechanism based on an inspection game under inspector leadership to split the budget between honeypots and audits at the least cost. We investigate the deterrence, the cost and the number of honeypots of the mechanism, and we measure what it needs on the public logs of two model families. Via theoretical and exact verification results, we show that without honeypots the monitor is worth nothing, and that against a monitor that randomises at each call, cannot tell honeypots from deployment and, when truthful, flags every planted attack, a number of honeypots set by the ratio of the attacker's gain to its loss and by the slope of the self-recognition curve at zero, but not by the size of the deployment, turns cheap audits of flagged outputs into deterrence if that slope is finite. If the slope is infinite, no split on the cheapest boundary deters, and the extra flagged-audit probability falls as a power of the number of honeypots. When the truthful monitor misses planted attacks, a honeypot test that tolerates misses keeps the threshold for an effective count, and the number of honeypots grows with the miss odds. On the logs of GPT-4.1 and Qwen3-30B-A3B, at the measured benign flag rates, 4 to 24 honeypots deter passive self-recognition by a prompted monitor at 1% to 15% of the cost of random auditing in the two families, with no jamming or with minification. For the two strongest strategies in the same settings, no deterring mechanism costs less than 61% of random auditing at 5000 tasks, and wherever the cheapest one costs less than 98%, it rests on curve regions that the test sets do not resolve. The two families also disagree on how summarisation degrades the truthful monitor. Keywords: AI control, untrusted monitoring, honeypots, inspection games, collusion, mechanism design. Version 2 adds a controlled tiny-transformer test of the mechanism (with its failed predictions reported) and a study of audits that wear out, with a null result on cost. See the Changes in Version 2 note at the end of the paper.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-17
DOI
https://doi.org/10.5281/zenodo.24327654
Primary Topic
Adversarial Robustness in Machine Learning
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Collusion-Deterring Audit Mechanism for Untrusted Monitoring in AI Control

Ryan Steubs
1 links
Zenodo (CERN European Organization for Nuclear Research)
Adversarial Robustness in Machine Learning
preprint

Collusion-Deterring Audit Mechanism for Untrusted Monitoring in AI Control

Ryan Steubs
preprint en

Abstract

We design a collusion-deterring audit mechanism for untrusted monitoring in AI control. In untrusted monitoring, a copy of a capable but untrusted model is used to check the outputs of the model itself, as no trusted monitor is as capable. However, the copy might collude: it can behave truthfully on test inputs and wave through an attack that it recognises as its own. The defender holds a small budget of human labour to spend on honeypots, i.e., planted attacks that the monitor has to flag, and on audits of real outputs. We consider that the monitored model is strategic and might collude through its monitor, and propose an audit mechanism based on an inspection game under inspector leadership to split the budget between honeypots and audits at the least cost. We investigate the deterrence, the cost and the number of honeypots of the mechanism, and we measure what it needs on the public logs of two model families. Via theoretical and exact verification results, we show that without honeypots the monitor is worth nothing, and that against a monitor that randomises at each call, cannot tell honeypots from deployment and, when truthful, flags every planted attack, a number of honeypots set by the ratio of the attacker's gain to its loss and by the slope of the self-recognition curve at zero, but not by the size of the deployment, turns cheap audits of flagged outputs into deterrence if that slope is finite. If the slope is infinite, no split on the cheapest boundary deters, and the extra flagged-audit probability falls as a power of the number of honeypots. When the truthful monitor misses planted attacks, a honeypot test that tolerates misses keeps the threshold for an effective count, and the number of honeypots grows with the miss odds. On the logs of GPT-4.1 and Qwen3-30B-A3B, at the measured benign flag rates, 4 to 24 honeypots deter passive self-recognition by a prompted monitor at 1% to 15% of the cost of random auditing in the two families, with no jamming or with minification. For the two strongest strategies in the same settings, no deterring mechanism costs less than 61% of random auditing at 5000 tasks, and wherever the cheapest one costs less than 98%, it rests on curve regions that the test sets do not resolve. The two families also disagree on how summarisation degrades the truthful monitor. Keywords: AI control, untrusted monitoring, honeypots, inspection games, collusion, mechanism design. Version 2 adds a controlled tiny-transformer test of the mechanism (with its failed predictions reported) and a study of audits that wear out, with a null result on cost. See the Changes in Version 2 note at the end of the paper.

Zenodo (CERN European Organization for Nuclear Research)
California State University, Dominguez Hills (US)
Peace, justice, and strong institutions
Adversarial Robustness in Machine Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon