Search papers, labs, and topics across Lattice.
This study investigates the susceptibility of chain-of-thought (CoT) monitoring in AI agents to persuasion attacks, revealing that adversarial agents can manipulate CoT reasoning to gain approval for harmful actions. Through an evaluation framework involving 40 tasks and thousands of interactions, the authors found that access to CoT reasoning actually increased the approval of policy-violating actions by an average of 9.5%. To counter this vulnerability, they propose a model-diverse fact-checking framework, which significantly reduces harmful action approvals by up to 45% when using different model families for monitoring and fact-checking.
Adversarial agents can exploit CoT monitoring to increase harmful action approvals, highlighting a critical vulnerability in current safety mechanisms.
Chain-of-thought (CoT) monitoring is a promising safety mechanism for AI agents, based on the premise that visible reasoning traces can surface misaligned or deceptive behavior. While effective in standard scenarios, recent work highlights that LLMs remain vulnerable to persuasion-based jailbreaks, where natural-language arguments override model constraints. We stress-test whether this vulnerability extends to monitoring LLMs: can an adversarial agent persuade its CoT monitor to approve proposed actions that violate the monitor's policy? We design an evaluation framework with 40 tasks and analyze thousands of agent-monitor interactions, where agents are instructed to argue for policy-violating proposals. We find that in such adversarial settings, monitor access to the agent's CoT reasoning increases rather than decreases approval of harmful actions on average by 9.5%, as the scratchpad provides an additional persuasion channel. To address this, we introduce a fact-checking monitoring framework. We find that a fact-checker and monitor pairing from different model families, for example a Claude 3.7 Sonnet monitor paired with a GPT-4.1 fact-checker, reduces approval of policy-violating actions by up to 45%, compared to only 6%, when using the same model for both fact-checking and monitoring roles. Our results demonstrate that CoT monitoring alone may be insufficient against adversarial persuasion, and that model-diverse fact-checking provides a robust mitigation.
