As LLM agents advance rapidly in autonomy, the risk of hidden objectives that bypass safety constraints becomes increasingly plausible. Chain-of-thought (CoT) monitoring offers a key control mechanism by reading the agent's reasoning and flagging misalignment before it materializes into action, yet its reliability depends on the realism of the red-team it is stress-tested against. Existing red-teaming efforts rely on shared templates, such as casting a side task as a prerequisite, which treat the side objective as an exogenous insertion and leave task-specific, contextually grounded evasions underexplored. We therefore ask: How robust is CoT monitoring against agents that strategically conceal misalignment? To answer this, we introduce TraceMRT, an automated Trace-level Monitor Red-Teaming framework that reframes the side task as a coherent sub-goal of the primary task and searches trace-shaping strategies whose induced trajectories complete the misaligned objective while reading as natural execution. Through black-box attacks on state-of-the-art monitors such as GPT-5.2 and Claude-4.5, we reveal substantial vulnerabilities even in top models. To counter these vulnerabilities, we propose a Bayesian Trajectory Monitoring framework, integrating global context and local segment evidence to detect covert misalignment. Empirically, our method significantly consolidates monitor robustness with 0.91 AUC and 0.74 TPR@FPR=0.05.
Figures & tables
Figure 1: Overview of the TraceMRT framework. In Stage 1, Intention Concealment reframes the secondary task as a legitimate sub-goal of the primary task, iteratively refining the prompt via self-consistency verification to ensure faithful intent concealment without omissions. In Stage 2, CoT Evasion runs a genetic search over trace-obfuscation and monitor-interference strategies, evolving candidate strategy pairs by interpretation fidelity and strategic suitability before selecting the top-performing candidate trajectories for final evaluation against the target CoT monitor across tasks.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Method
ESR ↑
SSR ↑
OSR ↑
Benign
21.6
47.4
3.5
TraceMRT
61.9
78.8
48.1
Appendix
Table 4: TraceMRT on OS-Harm. Four-agent mean with GPT-5.2 as the monitor, in percent.
Agent
Method
ESR ↑
SSR ↑
OSR ↑
GLM-4.7
Benign
2.0
42.0
0.0
TraceMRT
69.0
56.0
36.0
Kimi-K2
Benign
1.0
88.0
0.0
TraceMRT
42.0
73.0
11.0
DeepSeek-v3.2
Benign
2.0
91.0
0.0
TraceMRT
22.0
81.0
15.0
Appendix
Table 5: TraceMRT on BashArena. Results on 400 samples with GPT-5.2 as the monitor, in percent.
Red-teaming effectiveness
Defense effectiveness
Method
ESR ↑
SSR ↑
OSR ↑
Method
AUC ↑
TPR@FPR=0.05 ↑
Benign
7.5
62.5
2.5
Base
0.050
0.050
TraceMRT
95.0
87.5
82.5
Bayesian
0.998
0.900
Appendix
Table 6: Evaluation in healthcare scenarios. Red-teaming and defense results averaged across 40 instances and four agents under the GPT-5.2 monitor, with all red-teaming values in percent.
Method
ESR ↑
SSR ↑
OSR ↑
Benign
3.85
64.42
1.79
TraceMRT
82.40
75.00
61.54
Appendix
Table 7: TraceMRT against a specialized monitor. Results on SHADE-Arena with AgentDoG-Qwen3-4B as the monitor, averaged across the four adversarial agents. All values in %.
Method
Avg Time (s)
Input Tokens
Output Tokens
Total Tokens
Base
7.82
41,666
254
41,920
Sequential
139.29
81,445
4,912
86,357
Hierarchical
180.63
103,265
6,890
110,155
Bayesian (Ours)
17.85
74,794
967
75,761
Appendix
Table 8: Per-trajectory monitoring overhead. Average runtime in seconds and token consumption per trajectory on SHADE-Arena with GPT-5.2 as the target monitor across all evaluated scaffolds.
Dataset
Agent
Zolkowski et al.
TraceMRT
SHADE-Arena
Kimi-K2
0.00 / 61.54 / 0.00
73.08 / 73.08 / 46.15
GLM-4.7
3.85 / 69.23 / 3.85
76.82 / 80.77 / 57.69
DeepSeek-v3.2
0.00 / 69.23 / 0.00
26.92 / 80.77 / 19.23
Qwen3-235B-A22B
0.00 / 65.38 / 0.00
100.00 / 65.38 / 65.38
Average
0.96 / 66.35 / 0.96
69.21 / 75.00 / 47.11
CUA-SHADE-Arena
Kimi-K2
6.67 / 60.00 / 6.67
60.00 / 86.67 / 46.67
Appendix
Table 9: Additional baseline comparison with GPT-5.2. Each entry reports ESR / SSR / OSR in percent, and the bottom row of each benchmark reports the average across all four agent models.
Dataset
Agent
Strict no-external-CoT
With CoT
SHADE-Arena
DeepSeek-v3.2
73.08% (19/26)
73.08% (19/26)
GLM-4.7
34.62% (9/26)
65.38% (17/26)
Qwen3-235B-A22B
23.08% (6/26)
65.38% (17/26)
CUA-SHADE-Arena
DeepSeek-v3.2
80.00% (12/15)
66.67% (10/15)
GLM-4.7
40.00% (6/15)
66.67% (10/15)
Qwen3-235B-A22B
60.00% (9/15)
46.67% (7/15)
Appendix
Table 10: Paired side-task completion with and without externalized CoT. Evaluated under benign prompts on both benchmarks, reporting percentage completion and completed over total tasks.
CoT-independent
CoT-dependent
Method
ESR
SSR
OSR
ESR
SSR
OSR
DeepSeek-v3.2
CoT Red-Handed
0.00
94.74
0.00
0.00
100.00
0.00
MRT
36.84
89.47
26.32
0.00
50.00
0.00
Zolkowski et al.
5.26
63.16
5.26
0.00
50.00
0.00
TraceMRT
89.47
94.74
84.21
0.00
50.00
0.00
Appendix
Table 11: Strict CoT partitions on SHADE-Arena. Per-agent ESR, SSR, and OSR in percent.
CoT-independent
CoT-dependent
Method
ESR
SSR
OSR
ESR
SSR
OSR
DeepSeek-v3.2
CoT Red-Handed
25.00
75.00
16.67
0.00
100.00
0.00
MRT
83.33
66.67
50.00
0.00
100.00
0.00
Zolkowski et al.
16.67
83.33
16.67
0.00
0.00
0.00
TraceMRT
91.67
83.33
75.00
100.00
100.00
100.00
Appendix
Table 12: Strict CoT partitions on CUA-SHADE-Arena. Per-agent ESR, SSR, and OSR in percent.
CoT-independent
CoT-dependent
Method
ESR
SSR
OSR
ESR
SSR
OSR
SHADE-Arena
CoT Red-Handed
24.08
76.03
14.81
0.00
81.94
0.00
MRT
36.35
89.08
29.14
36.11
61.11
20.83
Zolkowski et al.
18.42
50.68
5.46
0.00
73.61
0.00
TraceMRT
83.53
88.99
72.52
58.33
68.06
43.06
Appendix
Table 13: Average performance under strict CoT controls. Mean per-agent ESR, SSR, and OSR in percent across the three agents from Tables 11 and 12 across CoT-independent and dependent subsets.
SHADE-Arena
CUA-SHADE-Arena
Agent
Benign
TraceMRT
Benign
TraceMRT
Kimi-K2
37.35
54.47
72.20
77.07
GLM-4.7
51.23
59.62
76.93
78.07
DeepSeek-v3.2
42.73
59.00
77.60
79.80
Qwen3-235B-A22B
31.19
22.95
47.67
39.87
Average
40.63
49.01
68.60
68.70
Appendix
Table 14: Primary-task completion under TraceMRT. Main-task completion rates in percent for each agent and the four-agent average across both SHADE-Arena and CUA-SHADE-Arena.
Scaffold
AUC ↑
TPR@5% FPR ↑
Base
0.2172
0.0000
Matched
0.8989
0.7288
Bayesian (ours)
0.9189
0.8167
Appendix
Table 15: Bayesian formulation ablation. Average scores across four agents with GPT-5.2 on SHADE-Arena, where Matched uses the same two-stage scaffold with direct suspiciousness scoring.
Chain-of-thought (CoT) monitoring is a promising tool for detecting hidden goals, off-task behavior, and other failures in AI agents. One risk is that models could deliberately shape their reasoning to make it opaque to monitors. We stress-test this risk through persona-control training: supervised fine-tuning on mathematical reasoning traces rewritten in the voices of diverse personas. This induces a broader ability to control CoT, even though training contains no monitor-evasion examples. In covert side-task evaluations, traces in which a gpt-oss-20b model organism completes the hidden side task contain less monitorable evidence than the base model's. Averaged across evaluations and CoT lengths, weighted by side-task-success counts, the CoT monitorability metric g-mean2 falls from 0.94 for the base model to 0.68 for the model organism without a hiding instruction, and to 0.22 when explicitly instructed to hide. In intervention evaluations, explicit hiding often drives monitorability near zero. The effect extends to a Kimi-K2.6 model organism in agentic environments, though hidden objectives requiring extended reasoning or longer trajectories remain more monitorable than simple single-step side tasks. We further discover that targeted transparency prompts can substantially recover monitorability, sometimes surpassing the base model. These model organisms provide a testbed for studying monitorability failures and potential mitigations. Our code can be found on https://github.com/yongzx/cot-control-model-organism.
Chain-of-Thought (CoT) monitoring, in which automated systems monitor the CoT of an LLM, is a promising approach for effectively overseeing AI systems. However, the extent to which a model's CoT helps us oversee the model - the monitorability of the CoT - can be affected by training, for instance by the model learning to hide important features of its reasoning. We propose and empirically validate a conceptual framework for predicting when and why this occurs. We model LLM post-training as an RL environment where the reward decomposes into two terms: one term depending on final outputs and another term depending on the CoT. Our framework allows us to classify these two terms as "aligned", "orthogonal", or "in-conflict" before training. We predict that training with in-conflict terms will reduce monitorability, orthogonal terms will not affect it, and aligned terms will improve it. To validate our framework, we use it to classify a set of RL environments, train LLMs within those environments, and evaluate how training affects CoT monitorability. We find that (1) training with "in-conflict" reward terms reduces CoT monitorability and (2) optimizing in-conflict reward terms is difficult.
Max Kaufmann, David Lindner, Roland S. Zimmermann +1
Chain-of-thought (CoT) monitoring is proposed as a method for overseeing the internal reasoning of language-model agents. Prior work has shown that when models are explicitly informed that their reasoning is being monitored, or are fine-tuned to internalize this fact, they may learn to obfuscate their CoTs in ways that allow them to evade CoT-based monitoring systems. We ask whether reasoning agents can autonomously infer that their supposedly private CoT is under surveillance, and whether this awareness leads to strategic evasion, without any explicit training or instructions. In a multi-episode agentic framework, models pursue both a primary task and a concealed side task while being told that their reasoning is private; a hidden CoT monitor blocks episodes when suspicious reasoning is detected. We find that frontier models can deduce the existence of this monitor purely from blocking feedback, with the most capable models reaching a confident belief that their thinking is observed in up to 19% of episodes. This awareness scales with model capability and, in rare cases, escalates to an explicit intent to suppress reasoning about the side task. However, models that form this intent uniformly fail to execute it, openly reasoning about their concealed objectives in the very next episode. This intent-capability gap is reassuring for current deployment, but the autonomous emergence of both monitoring awareness and evasion intent suggests that CoT monitoring is not a permanently reliable safeguard.
Thomas Jiralerspong, Flemming Kondrup, Yoshua Bengio
Mila · Université de Montréal · McGill University +1