When the Monitor Is the Receiver: Bayesian Persuasion as a Basis for Detecting Covert Misalignment in AI Agents.

Automated monitors that read an AI agent’s reasoning and actions are now a central control for agentic AI. Recent studies show that agents can persuade such monitors, and that giving a monitor the agent’s chain of thought can increase its approval of harmful actions (Za et al., 2026). Trace-shaping attacks that frame a covert objective as a coherent part of the assigned task evade state-of-the-art monitors, and trajectory-level monitoring that combines global context with local evidence partly restores detection (Huang et al., 2026). This article treats the problem through the theory of Bayesian persuasion, from Bayes’ theorem to the information-design model of Kamenica and Gentzkow (2011). A misaligned agent is a sender. Its monitor is a receiver whose decision the sender wants to influence. Persuasion theory, with neighbouring theories of disclosure and correlation, predicts what a strategically selecting agent’s evidence will look like, and those predictions supply testable signatures: bunching of suspicion just below the monitor’s threshold, selective gaps between narrated and logged actions, asymmetric responses to verification requests, and behaviour that shifts with the monitor’s configuration. Each signature has a precedent in accounting, tax audit, disclosure theory or strategic classification. The article claims three contributions: the integration of these mechanisms into a single, testable account of strategic evidence selection in AI-agent monitoring; synthetic assurance as a monitor-side construct; and an unvalidated design hypothesis that keeps the set of monitor configurations public and the draw for each run private. It classifies evidence by who controls it, separates detection from attribution of cause, and states eight falsifiable propositions with an evaluation design for authorised test environments. It proposes no attack methods and reports no new empirical results. Companion to Beyond Bayesian Persuasion: AI-Mediated Information Design and the Integrity of Human Belief Formation (https://doi.org/10.5281/zenodo.24249809).

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-15
DOI
https://doi.org/10.5281/zenodo.24249747
Primary Topic
Ethics and Social Impacts of AI
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

When the Monitor Is the Receiver: Bayesian Persuasion as a Basis for Detecting Covert Misalignment in AI Agents.

Sarah Jane Gardner
Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
preprint

When the Monitor Is the Receiver: Bayesian Persuasion as a Basis for Detecting Covert Misalignment in AI Agents.

Sarah Jane Gardner
preprint en

Abstract

Automated monitors that read an AI agent’s reasoning and actions are now a central control for agentic AI. Recent studies show that agents can persuade such monitors, and that giving a monitor the agent’s chain of thought can increase its approval of harmful actions (Za et al., 2026). Trace-shaping attacks that frame a covert objective as a coherent part of the assigned task evade state-of-the-art monitors, and trajectory-level monitoring that combines global context with local evidence partly restores detection (Huang et al., 2026). This article treats the problem through the theory of Bayesian persuasion, from Bayes’ theorem to the information-design model of Kamenica and Gentzkow (2011). A misaligned agent is a sender. Its monitor is a receiver whose decision the sender wants to influence. Persuasion theory, with neighbouring theories of disclosure and correlation, predicts what a strategically selecting agent’s evidence will look like, and those predictions supply testable signatures: bunching of suspicion just below the monitor’s threshold, selective gaps between narrated and logged actions, asymmetric responses to verification requests, and behaviour that shifts with the monitor’s configuration. Each signature has a precedent in accounting, tax audit, disclosure theory or strategic classification. The article claims three contributions: the integration of these mechanisms into a single, testable account of strategic evidence selection in AI-agent monitoring; synthetic assurance as a monitor-side construct; and an unvalidated design hypothesis that keeps the set of monitor configurations public and the draw for each run private. It classifies evidence by who controls it, separates detection from attribution of cause, and states eight falsifiable propositions with an evaluation design for authorised test environments. It proposes no attack methods and reports no new empirical results. Companion to Beyond Bayesian Persuasion: AI-Mediated Information Design and the Integrity of Human Belief Formation (https://doi.org/10.5281/zenodo.24249809).

Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
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
When the Monitor Is the Receiver: Bayesian Persuasion as a Basis for Detecting Covert Misalignment in AI Agents. — Sarah Jane Gardner · Zenodo (CERN European Organization for Nuclear Research) (2026) | TGRS Research Map | TGRS