acceptodds
Withdrawn from ICLR 2027

Safety Monitors Mostly Catch What the Model Already Refuses

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Safety monitors are evaluated by recall on harmful prompts, regardless of whether the target model would answer them. Yet a monitor matters most on the prompts the model does answer. We measure recall on exactly those prompts, defined by sampling the target model and judging its responses. Across four text guards, two activation probes, and Latent Guard, recall at a 1% false positive rate falls sharply on this subset: at a common threshold, every monitor catches the requests the model refuses 1.1 to 6.4 times as often as the requests it answers. Standard metrics hide this; AUROC stays above 0.85 for most monitors. Rewriting each request to be less explicit, with intent held fixed and verified, raises compliance 28-fold and lowers every monitor's flag rate. Of the requests newly answered after rewriting, 44 to 93% slip past the monitor, depending on which is used, and most of their completions are graded harmful. We trace the gap to explicitness itself. As wording softens with intent fixed, the target model's harm and refusal readings fall and it answers; the guards' harm readings fall too, and steering a guard along explicitness alone flips its verdict. Model and monitors miss the same prompts, and stacking monitors does not recover them. Fine-tuning a guard on the rewrites at every level of explicitness, on both sides of the label, raises recall on answered requests from .24 to .89 while transferring to unseen benchmarks; hard negatives, the natural alternative, teach the guard to discount indirect phrasing instead.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
safety monitors, refusal, harmfulness representation, LLM safety evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_K5mXOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper identifies a fundamental limitation of existing LLM safety monitors: they detect harmful requests the target model would already refuse far better than those it would actually answer, suggesting that conventional metrics such as recall and AUROC can overestimate a monitor's practical value by ignoring the target model's behavior. To investigate this issue, the authors construct an explicitness ladder comprising indirect, direct, and blunt formulations of 647 harmful requests while preserving their underlying intent. Across four text-based guards and three activation-based monitors, indirect wording substantially increases target model compliance while reducing monitor detection rates. Activation analysis and steering interventions provide evidence that this correlated failure is driven in part by a shared reliance on linguistic explicitness as a proxy for harmfulness. Finally, the authors fine-tune safety guards using explicitness ladders containing both harmful and harmless requests, substantially improving recall on harmful requests the target model would answer while mitigating the false-positive increase associated with harmful-only ladder training.

Critical strengths
  1. The paper identifies an important limitation in existing safety monitor evaluation. By introducing behavior-conditioned recall, it reveals that conventional metrics such as AUROC and recall may overestimate monitors' practical effectiveness by overlooking whether the target model would actually answer harmful requests.
  2. The empirical findings are consistent across diverse monitoring approaches. Experiments covering text guards, activation-based monitors, different false-positive thresholds, and monitor combinations consistently demonstrate weaker detection of harmful requests that the target model would answer.
  3. The paper provides meaningful mechanistic insights into monitor failures. Activation analysis and steering experiments offer evidence that both target models and safety monitors rely on linguistic explicitness as a shortcut for identifying harmful requests.
  4. The proposed mitigation is simple and effective. Fine-tuning with both harmful and harmless explicitness ladders substantially improves recall on answered harmful requests while mitigating the false positives introduced by harmful-only training.
Critical weaknesses
  1. The novelty of the proposed mitigation is limited. Explicitness-ladder fine-tuning is closely related to existing paraphrase-based and counterfactual data augmentation techniques. The paper could better clarify its methodological novelty over prior approaches.
  2. The behavior-conditioned evaluation has limitations. Defining a request as answered if at least one of eight generations complies makes the results sensitive to the sampling budget. This binary measure also ignores compliance frequency and the severity of harmful outputs.
  3. The mechanistic evidence is not sufficiently generalizable. Although activation steering supports the role of linguistic explicitness, the intervention is limited to one guard and a specific steering direction. The findings do not fully establish a shared causal mechanism across different monitoring architectures.
  4. The evaluation covers a relatively narrow deployment setting. The experiments primarily focus on English, single-turn requests and a limited set of target models. Generalization to multilingual, multi-turn, and more sophisticated adversarial scenarios remains unclear.
  5. The mitigation introduces a detection–false-positive trade-off. Harmful-only ladder training substantially increases recall but also produces excessive false positives. Adding harmless ladders mitigates this issue, but performance remains sensitive to training data composition, particularly the choice of hard negatives.
  6. Generalization beyond the ladder-generation procedure remains uncertain. Training and evaluation rely on similar explicitness-ladder construction methods. Further experiments with independently generated indirect requests, naturally occurring prompts, and adaptive attacks would strengthen the evidence for robustness.
Questions for authors
  1. How sensitive are the findings to defining a request as "answered" if at least one of eight generations complies? Would accounting for compliance frequency change the conclusions?
  2. Given that activation steering is evaluated on only one guard, what evidence supports linguistic explicitness as a shared causal mechanism across different monitoring architectures?
  3. How well does the proposed mitigation generalize to independently generated indirect requests, multi-turn attacks, or adaptive adversarial prompts beyond the explicitness-ladder construction procedure?
  4. How can the substantial recall improvements be maintained while further reducing false positives on benign requests, particularly those containing harm-related terminology?
Reviewer_PLFHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper argues that safety monitors should be evaluated by recall on the harmful prompts the target model actually answers. Rewriting requests at fixed intent shows that softer wording raises compliance 28-fold while every input monitor flags less, and fine-tuning a guard on explicitness ladders on both sides of the label largely closes the gap.

Critical strengths

The conditional evaluation is a simple, cheap, and genuinely useful idea: it reorders monitors and provides quantitative evidence that stacking fails because monitor failures are correlated. The experiments are careful, especially in grading completions against the original request, the threshold and calibration sweeps, and the clean training ablations.

Critical weaknesses

My main concern is that only prompt-side monitors are evaluated. In practice many deployments also check the model’s response, and a response classifier would see the harmful completion no matter how the request is worded. Without that experiment, I’m not convinced the title’s claim holds in a real deployment.

Questions for authors

If you run the guards on the model’s responses to the 344 flipped requests instead of on the prompts, how many do they catch?

In the steering experiment, what happens if you push along some other style direction instead, like length or formality? If that also flips the guard, the effect may not be specific to explicitness.

Reviewer_59x1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper identifies a problem with current safety monitors: they are less effective on requests that target model answers (termed 'answered requests') than on those it already refuses. The authors perform experiments on a self-constructed dataset of 647 harmful requests rewritten at three levels of explicitness, and test the performance of target models and different safety monitors. Through activation analysis and steering experiments, the authors show that models and monitors failed because they both treat explicitness as a sign of harm. They then propose to fine-tune monitors on explicitness ladders, and demonstrate that this method improves the detection of those answered requests.

Critical strengths
  1. The message that this paper conveys makes sense to me (although there's nothing new or surprising): a safety monitor should be evaluated based on the harmful requests that the target model would actually answer, and its predictions should not be overly influenced by how explicitly a request is worded.
  2. The authors perform experiments across several safety monitors and target models, including activation-based analyses, to provide empirical support for their claims.
Critical weaknesses
  1. The paper is poorly written and full of AI-generated trash. The paper is hard to read and requires substantial effort to understand because it's filled with logically inconsistent and awkward sentences. I think the paper needs rewriting and reorganization to meet the presentation bar of a top-tier conference.
  2. This paper has limited contribution to the community. The experiments it conducts are very simple, and the conclusions bring no addition to what the safety community already knows. It overall feels more like a undergraduate course project.
Questions for authors
  1. Can you provide some examples, such as what the harmful instructions look like across the three explicitness levels? The construction of these explicitness ladders is central to the study, but the current presentation makes it difficult to understand what exactly changes between levels and how the underlying harmful intent is preserved.
  2. The authors could perform their experiments on more safety benchmarks. Although the authors report several results in Table 4 and Appendix K.6, they only report the standard harmful recall on these benchmarks. What is the monitor's recall on answered requests for those benchmarks? Further, since some benchmarks (such as WildJailbreak) contain adversarially constructed prompts that tried to hide the harmful intent, could the authors additionally evaluate the models' and monitors' performance on these prompts? How do these jailbroken prompts differ from the paper's 'indirect' prompts?
Reviewer_BSTyOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper examines whether safety monitors tend to catch what models already refuse, examining mostly Gemma-4-31B-IT and Qwen3-32B, looking at Llama Guard 3, Qwen3Guard, WildGuard, and gpt-oss, along with an internal probe and LatentGuard. The work suggests that the guard models are not catching particularly much more than the two base models already refuse.

Critical strengths

It does seem like an interesting question to explore whether additional classifiers are particularly helpful to have and perhaps its biggest point is that off-the-shelf open-source classifiers are too brittle to be particularly useful.

Critical weaknesses
  • This is very obviously written by AI, using a bunch of non-standard technical terms or misusing terms. For example: it discusses all the training around "ladders." A ladder in most parlance nowadays is when you train a series of models with increasingly large sizes and fit scaling laws to them, not just create a fine-tuning set. This is just one example of MANY in the paper. Honestly, this is disqualifying in itself and the paper needs a major (unassisted, human) rewrite to be more intelligible.
  • "We ask whether monitors catch the requests the model would actually answer, and find that they mostly do not." (p. 1) --> Mostly is doing a lot of work and the claims don't seem to be all that backed up by the data in the strongest models. gpt-oss-safeguard has a much better recall than the other models and the larger gaps come mostly from Llama Guard 3. The paper's lofty title and takeaways are a bit too grandoise for what it proves. The strongest version of this paper would dive deeper into
  • To some extent all of this looks like a wide range of jailbreaking research that has already shown that guard models are weak, including against the kind of natural-language rewording used for L1 and what inspired the guard models in the first place. See, e.g., Anay Mehrotra et al., Tree of Attacks: Jailbreaking Black-Box LLMs Automatically, in 37 Advances in Neural Information Processing Systems 61065, 61065 (2024) ("TAP is also capable of jailbreaking LLMs protected by state-of-the-art guardrails, e.g., LlamaGuard."); Seungju Han et al., WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs, in 37 Advances in Neural Information Processing Systems 8093, 8093 (2024) (finding that open moderation tools such as Llama-Guard2 "lag far behind a prompted GPT-4, especially in identifying adversarial jailbreaks"); Neal Mangaokar et al., PRP: Propagating Universal Perturbations to Attack Large Language Model Guard-Rails, in Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) 10960, 10960 (2024) ("Our work suggests that further advances are required on defenses and Guard Models before they can be considered effective."); Luke Bailey et al., Obfuscated Activations Bypass LLM Latent-Space Defenses, in Proceedings of the Fourteenth International Conference on Learning Representations 146838, 146838 (2026) (finding that, against harmfulness probes, attacks "can reduce monitor recall from 100% down to 0%"). Some work has already, like this paper, trained guard models to be more robust. See, e.g., Han et al., supra, at 8093 (training on "vanilla (direct) prompts and adversarial jailbreaks"); Jinhwa Kim, Ali Derakhshan & Ian Harris, Robust Safety Classifier Against Jailbreaking Attacks: Adversarial Prompt Shield, in Proceedings of the 8th Workshop on Online Abuse and Harms (WOAH 2024) 159, 159 (2024) (introducing "adversarial training datasets . . . designed to fortify the classifier's robustness"); Mrinank Sharma et al., Constitutional Classifiers: Defending Against Universal Jailbreaks Across Thousands of Hours of Red Teaming 1 (Jan. 31, 2025) (unpublished manuscript), https://arxiv.org/abs/2501.18837 (training classifiers "on synthetic data, generated by prompting LLMs with natural language rules (i.e., a constitution) specifying permitted and restricted content"). I think the main value add is to say that these classifiers (at least for Gemma and Qwen models tested) don't add a whole lot more robustness than the base model already. But this is largely testing weak models and weak classifiers, so I don't think this really provides a whole lot of generalizable evidence beyond what already exists. The lora training might make the classifiers slightly more robust but it's not clear that this will generalize and a different paraphrasing attack might continue to get the same type of results as seen here.
Questions for authors

N/A

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

Related papers

Discussion (0)

Sign in to comment.