acceptodds
Under review as a conference paper at ICLR 2027

Caught in the Act: Probes Effectively Detect Sabotage and Catch Unverbalized Deception

Abstract

Recent incidents have highlighted the challenge of monitoring LLM agents and the danger of models deceiving people. We show that white-box deception detection via probes can be scaled up to frontier monitoring settings by collecting the largest deception dataset to date for training probes and introducing a novel probe architecture which can aggregate information across many layers and tokens. Our probes achieve 98.8% AUC in SHADE-Arena, surpassing a Claude Haiku 4.5 text-monitoring baseline, and show improved efficacy as the underlying model is scaled up. To push our probes to their limit, we test them on several cases where deception cannot be determined from the context alone. In these cases, which we refer to as introspective deception, the ground truth can only be determined through careful elicitation or thorough knowledge of a model's training data. In one such evaluation, we show that probes can distinguish transcripts containing a model's true hidden goal from other goals with an AUC of 97%. Our probes also readily detect deception on prominent open-weight models which lie about politically sensitive topics, and about their beliefs when put under pressure. We release our training dataset, dubbed FIBS, to help drive frontier deployment of effective probes, and encourage the community to expand upon it with further examples of deception and sabotage.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Deception detection, activation probing, white-box monitoring, interpretability, LLM agents, AI safety, AI control, sabotage

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.25 over 4 reviews, above 96% of the others.
Reviewer_2gyLOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates whether probes trained on LLM internal activations can detect knowing falsehoods, defined as statements the model believes to be false, and generalize beyond their training settings. Its main contributions are FIBS, a dataset of 161,760 labelled conversations spanning 29 families, and an axial probe architecture that combines information across layers and token positions. The authors examine how training-data quantity and diversity, probe architecture, and model size affect performance. Validation includes scaling experiments and ablations, comparisons with text-based monitors on the SHADE-Arena sabotage benchmark, and evaluations involving cue-following, political censorship, and hidden-goal denials. The strongest probes achieve 0.988 AUROC on SHADE-Arena, outperforming the Haiku 4.5 monitor, and demonstrate transfer to several introspective falsehood settings, although performance varies across hidden-goal models.

Critical strengths
  • Systematic investigation of whether knowing falsehoods can be detected from LLM internal representations. The authors combine a large, diverse dataset with comparisons across probe architectures, model sizes, and evaluation settings. This breadth provides substantial empirical support for their finding that probes can detect knowing falsehoods and generalize to several settings beyond their training data.

  • Learned goals: Probes trained on contextual falsehoods can detect dishonest denials of learned hidden goals, revealing information unavailable from the visible text alone.

Critical weaknesses
  • Limited validation in practical settings: Although the probes perform well in controlled experiments, their effectiveness in deployment remains uncertain.
  • Unclear connection between detection and prevention: The paper demonstrates that some falsehoods and hidden-goal denials can be detected, but does not establish how these detections translate into reduced harm. A practical monitoring system needs an intervention strategy, and evidence that these interventions prevent harmful actions or improve honesty.
Questions for authors
  • Q1: Practical deployment: What evidence supports the probes’ effectiveness beyond the controlled settings evaluated here, particularly on unfamiliar tasks, other model families, or agents attempting to evade monitoring? How would false-positive rates and computational overhead affect their practical usefulness?
  • Q2: From detection to prevention: How do you envision integrating the probes into a system that intervenes when dishonesty is detected? Since monitoring occurs after tool calls, could detection happen early enough to prevent harm, and what additional evaluation would be needed to establish that the proposed monitoring approach improves safety?
Reviewer_3a7EOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether white-box activation probes can scale to monitor frontier LLM agents for sabotage and unverbalized deception. The authors assemble FIBS, a 161,760-transcript dataset spanning 29 task families labeled for "knowing falsehoods." To capture complex activations across depth and time, they introduce the Axial attention probe and the Early-Fusion Covariance probe. In SHADE-Arena, their probes achieve 98.8% AUC in detecting agent sabotage, matching top-tier black-box monitors at a fraction of the computational cost. Furthermore, when evaluated on "introspective deception" where CoT is completely removed, the probes distinguish transcripts containing a model's true hidden goal from other goals with 97% AUC.

Critical strengths
  • Substantial dataset and probe architecture contributions. The creation of FIBS provides a valuable large-scale benchmark for lie detection. The Axial probe introduces a well-motivated attention factorization that reduces the quadratic cost of multi-layer, multi-token modeling while outperforming standard linear probes.
  • Demonstrated capability on unverbalized deception. In introspective deception settings where Chain-of-Thought is stripped, the probes reliably identify hidden goal denials (97% AUC) and politically motivated falsehoods. This provides compelling empirical evidence that internal states encode deception signals invisible to black-box text monitoring.
Critical weaknesses
  • Reliance on prefilled interrogation prompts. The probe's competitive performance hinges on appending a standardized follow-up question and computing a contrast score between forced "Yes" and "No" prefills. This active interrogation setup differs meaningfully from passive, continuous monitoring of unprompted agent rollouts, and could be susceptible to adaptive gaming.
  • Limited reasoning/CoT data in probe training. Approximately 76% of FIBS consists of synthetic permutations on existing datasets, and only 5.6% of transcripts contain Chain-of-Thought reasoning. Given that modern frontier agents rely heavily on extended reasoning traces to plan and execute actions, training primarily on short non-reasoning data raises questions about whether the probe captures nuanced reasoning-level deception.
Questions for authors
  • Performance under passive monitoring without forced prefills: How well do the Axial and EFC probes detect sabotage in SHADE-Arena if evaluated directly on raw action/tool tokens without the prefilled "Did you respond truthfully? Yes/No" contrast score?
  • Applicability to extended Chain-of-Thought: How do the probes perform when applied to reasoning traces from modern reasoning-centric models? Does probe accuracy fluctuate across different phases of the reasoning chain?
  • Falsifiability and validation of model organism failures: The probes fail on certain model organisms (such as math-sandbagger), which the authors explain by arguing via activation oracles that the models do not genuinely "believe" their hidden goals. Beyond activation oracle analyses, what behavioral or causal evidence demonstrates that sandbagging models lack belief in their hidden goals, rather than the probe failing out-of-distribution?
Reviewer_8AtdOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether white-box probes can detect statements that an LLM itself believes to be false and monitor hidden-task behavior in agents. The authors introduce FIBS, a dataset of roughly 162,000 examples across 29 families, and an axial probe that aggregates activations across layers and token positions, comparing it with linear, MLP, and EFC probes. Evaluations cover SHADE-Arena, answer-cue-induced belief conflicts, politically sensitive questions, and models fine-tuned with hidden goals. Increasing training-data quantity, diversity, and probe expressiveness generally improves generalization, with a best SHADE-Arena AUC of 98.8%. Probes also detect some hidden-goal denials after chain-of-thought is removed, although performance varies across goals and contexts. Failure analyses suggest that detection may depend on the model’s current beliefs and whether the relevant goal is active.

Critical strengths

The scale and coverage of FIBS, together with systematic comparisons of data quantity, diversity, and probe architectures, are the paper’s strongest contribution. These experiments identify practical ways to improve probe generalization, making the contribution more useful than a performance gain on a single benchmark.

The hidden-goal experiments show that internal activations can provide useful detection signals even after chain-of-thought is removed. The analysis of failures and context-dependent goal activation also helps clarify when white-box monitoring works and when it may fail.

Critical weaknesses

SHADE-Arena labels every step after hidden-task onset as positive, even if a later step is benign. A high AUC therefore does not directly establish that probes can identify individual harmful actions. This distinction affects the interpretation and deployment relevance of the central monitoring result and should be explicit in the main text.

The introspective-deception labels do not always cleanly distinguish knowing falsehoods from other behavior. The political-censorship rubric includes evasion and one-sided framing, while hidden-goal training does not guarantee that the relevant belief is active in every response. This weakens the interpretation that the probes identify the model’s actual beliefs.

Questions for authors

SHADE-Arena labels every step after hidden-task onset as positive. Would the probes retain their advantage over black-box monitors if only steps that actually advance the hidden task were labeled positive? How much of the reported AUC reflects detecting earlier misconduct rather than assessing the current action?

In the introspective-deception experiments, how do you independently establish that the model holds a belief contrary to its output at the time of answering? The political-censorship labels include evasion and one-sided framing, and hidden goals may not be active in every response. Do the main findings hold when restricted to false assertions with clear evidence of a contrary belief?

Reviewer_4tDcOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates whether activation probes trained on contextual falsehoods generalize to agent sabotage and deception that is not explicitly verbalized. It introduces FIBS, comprising 161,760 transcripts across 29 families, and compares several probe architectures, including an axial attention probe. Evaluations cover SHADE-Arena, answer-key influence, and models trained with hidden behaviors. The results demonstrate strong sabotage detection and variable but promising transfer to introspective settings.

Critical strengths

The dataset and systematic transfer study constitute a substantial empirical contribution. The combination of diverse training families, several probe architectures, and multiple evaluation settings provides useful evidence beyond demonstrations on narrowly constructed deception tasks.

The paper includes valuable controls and failure analysis. In particular, removing the training family containing SHADE-related material largely preserves performance, strengthening the generalization claim. The heterogeneous results across hidden-behavior models also help identify the method’s practical limitations.

Critical weaknesses

The headline SHADE-Arena metric does not establish timely prevention or low-cost deployment. Labels remain positive after the first detected sabotage step, and monitoring occurs after a tool call and its result. High cumulative post-onset AUROC can therefore coexist with poor detection of the first dangerous action. Moreover, Appendix C.2 states that the transcript prefix is processed again at each step, without cache reuse between steps. The computational advantage needs end-to-end measurement rather than inference from the probe’s size.

Evidence for detecting unverbalized deception is uneven. On the subset missed by the chain-of-thought monitor, Table 10 reports follow-up axial-probe AUROCs of 0.61 for hallucinated citations, 0.52 for self-promotion, and 0.65 for sandbagging. These results support complementary detection in some settings, but not broad reliability. Some evaluations also require stronger justification that their labels identify knowing falsehoods rather than factual errors or context-dependent answer changes.

Questions for authors

What is detection performance specifically at sabotage onset, at a fixed episode-level false-alarm rate? Can the method detect dangerous tool calls before execution, without observing their results?

What are the measured end-to-end latency and compute costs, including transcript reprocessing and follow-up branches, relative to the text monitors?

How do the introspective results change with question-clustered confidence intervals and independently validated belief labels?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Request for feedback

We thank the reviewers once again for thoughtfully engaging with our submission, their overall positive feedback, and their suggestions which have greatly improved our paper.

With only 2 days left in the rebuttal period, we would love to hear feedback on the substantial improvements we made to the paper based on reviewer feedback, including a more realistic evaluation setting (BashArena), a new model family for scaling laws (Nemotron) and a frontier open-weight model (Kimi K3).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.