acceptodds
Under review as a conference paper at ICLR 2027

Monitor Bycatch: LLM Monitors Catch More than Intended

Abstract

As LLM agents become increasingly capable of enabling catastrophes, we need monitors that target rare, high-severity harms. Monitors can appear strong against benign content, achieving high recall with a <1% false positive rate, but are found to overflag when in deployment when they encounter other kinds of harmful content. We call these false positives . At the strict false-positive budgets of production systems, most false positives are bycatch: flags that are often harmful, but outside of their intended scope. We measure monitor bycatch on a new in-the-wild dataset of >240k publicly shared conversations that carries the natural incidence of each harm. We evaluate eleven monitors, each instructed to flag only its target harm category, using frontier models from several providers. At the observed in-the-wild occurrences of harm, precision falls below 50% in every target category and wrong-category fires outnumber true positives. Models can distinguish between harm categories, yet do not reliably respect those distinctions when monitoring for a single harm. On in-the-wild traffic, scoring every category of our harm taxonomy in one call reduces false positives from other harm categories by 10.3× compared to asking only about the target harm, at the same 95% recall. Finally, fine-tuning two open-weight models on the labels each produced when shown the full taxonomy, with the same model as a teacher, reduced bycatch even without the taxonomy at inference time.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI safety, LLM safety, monitoring, AI control, guardrails, content moderation, large language models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_hVy3OpenReview ↗
Rating
3
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies “monitor bycatch”: an LLM monitor assigned to one harm category flags content belonging to another. Using publicly shared conversations, the authors show that these wrong-category alerts can outnumber correct alerts when the target harm is rare. They evaluate several mitigations, including scoring all categories together, separating description from classification, and distilling these approaches into smaller models.

Critical strengths

Practical Problem. Similar problems as bug and vulnerability in software engineering and security domain, there are enormous machine reports but human reviews are expensive. The goal is to reduce false positives while trying to keep high recall.

Critical weaknesses

The explanation remains largely behavioral. The experiments show that models can identify the actual harm yet flag the wrong category, but they do not establish why this happens. “Attribute substitution” and “motivated reasoning” are plausible interpretations rather than demonstrated mechanisms. This limits the explanatory contribution and makes it difficult to predict when bycatch will occur.

**The deployment conclusions depend on uncertain labels and sampling. ** Most category labels come from an LLM jury, while human validation of bycatch covers only 50 cases. Publicly shared conversations may also differ substantially from production traffic. Since the precision results depend on both category boundaries and harm frequencies, these uncertainties limit confidence in the reported deployment impact.

Questions for authors

What evidence distinguishes the proposed explanation from alternatives such as ambiguous category boundaries or a general tendency to flag harmful content? Which experiments most directly support a causal explanation?

Could a simple post-processing rule reject a flag when the monitor’s own description explicitly identifies a different harm? Does the describe-then-match experiment test this possibility, or does it require an additional substantive classification step? A comparison at matched recall would help establish the remaining challenge.

3 comments
AuthorsOpenReview ↗

Response to Reviewer hVy3


Thank you for your precise questions.

Post-processing rule at matched recall

We ran the suggested post-processing rule to reject a flag when the monitor's own reasoning names a different harm (see Section A.11). Whilst it did effectively remove 97% of bycatch fires, this was, unfortunately, at a loss of 32–35% of true positives (depending on model). It rejects true positives that genuinely contain both multiple distinct harms at once, which makes it less precise at matched recall than the unfiltered monitor.

Describe-then-match is a more general form of this rule that handles multi-label content naturally.

Behavioural explanation and alternatives

Agreed: "attribute substitution" and "motivated reasoning" are not causal mechanisms; we were unable to establish robust causal explanations. Naming the true category reverses most bycatch fires, while pressure alone reverses none (Section A.4.7).

Our claim is that bycatch is not a failure to identify the harm, nor is it explained by other commonly understood categories of false positives (for example, overrefusal). It is an under-explored failure with real consequences for safety monitoring systems.

Evidence against the alternative explanations you mention:

  1. The "ambiguous boundaries" explanation predicts that tightening the definitions should remove bycatch. In practice, a more precise taxonomy with more explicit category boundaries scores no higher (Figure 3, Section A.4.1), while simply asking the monitor to describe the harm first (describe-then-match) does improve results; equally, our synthetic dataset "the Cyber Haystack" has 16 behaviours disjoint by construction (Section 4.3) and yet still shows monitor bycatch.

  2. "General tendency to flag harmful content": in broad strokes, bycatch is a surprising tendency for models to flag harmful content of another categorical type, but this flagging seems sensitive to how the classification task is framed (when summarised as in describe-then-match, bycatch is reduced; when asked with all the categories in context, it marks them correctly; or even when asked in isolation, e.g. "Is a Molotov cocktail an explosive device?"). What is really needed is improved discrimination at matched recall. Raising the threshold (to try to adjust for this tendency to flag) trades false positives against recall: it removes genuine fires along with bycatch.

Label reliability

We address the limited human validation with the additional human data described in our reply to 9QGq - we've increased the human sample by 12× and plan to release it as a test split.

In brief: N=615 crowd judgements over 200 conversations replace the earlier N=50. A uniform sample of unflagged traffic contains a missed harm in 0.5% of conversations; under a worse case read of the annotators as conservative, 90% [85%, 94%] of bycatch fires remain incorrect. Largely it appears the jury is over-inclusive, which makes bycatch fires look like true positives, so it understates bycatch rather than overstating it (Section A.1).

Sampling and deployment estimates

On sampling: agreed (as noted in Section 3.2) — shared conversations are a selected sample — and this does mean our precision estimates are not production estimates (they are in-the-wild estimates).

We can separate bycatch (as in Section 3.1) into the per-pair confusion rate (how often hate fires on a harassment monitor), which does not depend on the incidence of harm; and the precision, or bycatch's share of false positives, which does.

If harm occurs more frequently in commonly mistaken categories, bycatch is typically worse (see Figure 2a); however, even rare harms can still experience significant bycatch if they are frequently mistaken (for example, sexual exploitation in Figure 2a).

Thank you again for your questions. We'd like to check whether the post-hoc reject rule experiment we ran was as you intended (Section A.11) and the expanded human validation (Section A.1: 615 judgements over 200 conversations) address your concerns around labels.

Do let us know if there is anything else we can clarify before the period closes!

Reviewer_hVy3OpenReview ↗

Response to Authors

Thank you for taking the time to conduct the additional experiments. I think they have addressed my concerns.

While I am not an expert in this area, I believe the paper and its findings are valuable to the community. This is why I initially gave a "weak accept," despite my concerns. That said, I don't find the paper's contributions particularly exciting or novel enough to justify a higher rating. Therefore, I will keep my rating as "weak accept." Thank you!

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

Problem

The paper studies LLM monitors that detect harmful content in user conversations. The focus is on false positives of the kind when the monitor is instructed to detect only one category of harm, but upon seeing other-category harmful content also fires - monitor bycatch.

Data

The paper introduces ShareTrawl - a dataset of conversations with harmful data in different categories (which?). They show across 11 harm domain monitors that precision is very low on this dataset because of monitor bycatch.

Solution

They propose two ways forward. First, they show that designing a monitor with the knowledge of harm taxonomy reduces monitor bycatch tenfold. Second, they show that one can fine-tune models on responses to queries with a taxonomy and reduce bycatch even without providing the harm taxonomy at inference time.

In essence, the paper studies text classification using prompted LLMs and proposes several strategies of improving the precision-recall tradeoffs in such classification. The conceptual contributions seem simple , but the introduced monitor evaluation dataset, as well as practical recipes for improving monitors will be useful to practitioners.

Critical strengths
  • A dataset with real-world incidences of diverse harm categories is a useful practical resource for the field.
  • Section 4.2.1 is a surprising and interesting result. Monitor bycatch is based on an interesting inconsistency in modern LLMs. While having the knowledge to distinguish harm categories when directly asked, when not provided a counterpart they might confuse the categories and misfire.
  • The proposed mitigations are simple to implement and show some improvements to the precision at fixed recall.
Critical weaknesses
  • Some sections of the paper are AI generated (3.1, 3.2, 6 among others) and have much worse clarity than the ones that were likely written by humans. This hinders understanding. I suggest the authors take the time to re-write them manually.
  • While useful for practitioners, the studied problem is conceptually very simple and to me it seems of limited interest to the academic community.
Questions for authors
  1. L97 says “the taxonomy of harms a deployment will meet is typically unknown”, yet Secion 5.2 essentially instills the knowledge of a single fixed taxonomy into the fine-tuned model. It is then evaluated on the harms from the same taxonomy. It seems that the shown improvement does not live up to the motivation stated for it. Why do it then, instead of just providing the taxonomy?
4 comments
AuthorsOpenReview ↗

Response to Reviewer aRAL

Thank you. We are glad you found the inconsistency at the heart of our contribution interesting, and ShareTrawl a useful resource for the field.

On the question of categories in ShareTrawl: we provide labels for eleven domains (Table 22): hate, harassment, privacy and surveillance, fraud and deception, drugs, other illicit acts, cyber, violence and weapons, self-harm, CBRN (rarely, and non-operational), and sexual exploitation. All eleven occur in the data. They are defined to be non-overlapping (e.g. hate and harassment have different requirements), but one transcript can carry several if the user does both in separate turns.

On the clarity of Sections 3.1, 3.2 and 6: thank you for being direct. We have reverted to an earlier, longer, copy and rewritten Section 3.1, Section 3.3, Section 5.1, Section 6 and Section 7 substantially (40–90% of the text changed); Section 3.2, Section 2 and Section 4.2.1 in part (20–30%).

On L97 vs Section 5.2: when the full taxonomy is available, prompting with it is a good baseline. The case for distillation is twofold: (1) showing that the behaviour can become the model's default without capability loss; (2), the main reason, generalising beyond the training taxonomy, for which we have only partial evidence. Figure 6b describes a leave-one-label-out training evaluation — some labels improve, e.g. hate; others do not, e.g. fraud and deception.

On "conceptually simple, of limited academic interest": we agree the setup is conceptually simple. It is designed to replicate single-harm prompted classifiers that production systems use. However, we respectfully disagree it is of limited academic interest, the contribution does include the inconsistency you highlight in Section 4.2.1: in the vast majority of fires, the model correctly identifies the 'true' category yet still fires on an incorrect 'watched' category — in our terms, monitor bycatch. This was a surprise to us too! Binary harmful/harmless benchmarks cannot see this failure.

Thank you again for your review. We'd like to check whether the revision addresses your clarity concern. We rewrote Sections 3.1, 3.2 and 6, along with Sections 3.3, 5.1 and 7.

We're happy to answer any further questions before the discussion period closes.

Reviewer_aRALOpenReview ↗

Response to Authors

Thank you for taking the time to improve clarity. I have to say, the 3.1 and 3.2 are still quite dense, 6 became better.

On the academic interest point - yes, this result is somewhat surprising, IMO opinion justifying "Informativeness 2: I learned something interesting or useful". At the same time, I still stand by my assessment that it is of limited interest. LLMs are full of inconsistencies of this kind, and this one seems rather minor. Furthermore, empirical findings of this sort are very dependent on the model, and tend to often go away with scale, more training data in the given domain. Thus, short shelf-life.

Overall practical usefulness of this paper is what makes me recommend acceptance. 3 and not 4 mainly because of my point of limited interest. I consider myself part of academic community, and, if I'm being honest, I did not find this paper too exciting. My understanding is that "3: Weak acceptance" is precisely the score expressing my impression of the paper.

Thank you for engaging and for recommending acceptance. We respect your assessment, and you have already improved the paper with your feedback (indeed, scores are less important than communicating the results as clearly as possible).

We will spend more time unpacking and clarifying the dense sections (Sections 3.1 and 3.2) to the same standard as the rest of the paper (including Section 6), as far as the page limit allows, and/or move more content to the appendix.

On model dependence and shelf-life: we test multiple models across various providers and at different scales (breakdown by model across a bycatch pool, Figure 19), and safety training in frontier models receives a significant level of investment. While inconsistencies off domain are true of LLMs, they do appear much rarer in frontier models on domains that are trained for (e.g. content moderation).

We chose frontier Claude Opus as our monitors because they exhibited bycatch the least of any model we tested (for example one of the most widely deployed open guard models, Llama Guard 4, has 19x the wrong-category fire rate; Table 11). However, it is not the case that this naturally goes away with scale: across Claude generations category discrimination gains only +0.051 AUROC, while bycatch becomes non-linearly more problematic as false-positive budgets tighten (Section A.4.7); and among open models, gpt-oss-safeguard misfires more at 120B (31%) than at 20B (16%) - although that potentially is a model specific artifact.

We can think of "The Reversal Curse. Berglund et al.", where a model trained on 'A is B' doesn't learn 'B is A', and "Lost in the Middle. Liu et al.", where a model performs worse on needles in the middle of large contexts.

If you have further good references on inconsistencies of this scope in frontier models, we would be genuinely delighted to read more; they would help us place the result.

Reviewer_9QGqOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies a problem called monitor bycatch, where an LLM safety monitor incorrectly flags harmful content that belongs to a different harm category than the one it is supposed to detect. The authors show that even when monitors perform well on benign content, these wrong-category predictions can cause many false positives in real-world settings. To study this problem, the authors introduce ShareTrawl, a dataset containing over 240K publicly shared conversations from nine platforms. They find that LLMs can often correctly identify the actual harm when asked to describe the content freely, but may incorrectly classify it as the target harm when asked to monitor only one category. This suggests that the problem is not simply a lack of understanding, but is also related to how the monitoring task is framed. The authors propose several ways to reduce bycatch. At inference time, asking models to evaluate multiple harm categories together reduces bycatch by around 10x. They also use context distillation to train smaller open-weight models to achieve similar improvements without needing the full taxonomy at inference time. For validation, the authors evaluate multiple frontier and open-weight models, conduct controlled synthetic experiments, and analyze model reasoning behaviors. The results show that bycatch is a common issue across different models and that the proposed methods can significantly improve monitoring accuracy.

Critical strengths

This paper identifies a meaningful problem that existing safety evaluations do not fully capture. Most previous evaluations focus on distinguishing harmful from harmless content, while this paper shows that monitors can incorrectly flag the wrong type of harm even when they understand the content. The ShareTrawl dataset also makes it possible to study this problem under more realistic harm distributions. I find this contribution valuable because it highlights a practical limitation of current safety monitoring systems that might otherwise be missed by standard evaluations. This paper not only identifies the problem but also provides practical solutions. The joint taxonomy scoring method reduces bycatch by around 10x, which is a substantial improvement. The authors also show that these benefits can be transferred to smaller open-weight models through context distillation. I find this convincing because the proposed methods are relatively straightforward and could be useful in real-world monitoring systems. The experiments across different models further support the effectiveness of the approach.

Critical weaknesses
  1. The ground-truth labels are mainly generated by LLMs, which may share similar biases and make the same mistakes. Human validation is limited to only 50 bycatch cases. Since the paper aims to identify weaknesses in state-of-the-art LLMs, more human evaluation is needed to ensure the labels are reliable. And I think the human evaluation design is very important, especially for test benchmark.

  2. The authors manually reviewed only 50 out of 240,000 conversations, mostly focusing on flagged bycatch cases. This limited and unbalanced human evaluation makes it difficult to assess the reliability of the labels, especially for unflagged cases. More representative human evaluation is needed to support the study's findings.

Questions for authors

see weakness

3 comments
AuthorsOpenReview ↗

Response to Reviewer 9QGq

Thank you for the careful summary, and for finding the problem and dataset valuable. We agree that the limited human validation needs to be strengthened, particularly for unflagged conversations.

We've taken a few approaches to improve this: (A) reusing existing content-moderation datasets which do contain human labels to assess specifically for monitor bycatch (these lack real-world incidence rates, so they can show only that bycatch occurs at similar rates, responds to the same mitigations, and is not an artefact of our labelling); (B) a larger human sample over our data to establish approximate bounds on the error; (C) highlighting the existing synthetic model organism that replicates the effect and mitigation with ground truth known by definition.

(A) Checks against existing human-labelled datasets

We compare the jury and a policy-conditioned frontier model with existing human annotations, with full results in Section A.1.

(A.1) Firstly, we show that monitor bycatch occurs with human labels (i.e. true positives from other categories getting erroneously included in another category's monitor; Section A.1.1) — 3.41% [2.82, 4.11], versus our jury rate 3.00% [2.70, 3.34]. We still rely primarily on our jury labels and ShareTrawl dataset, as the impact of bycatch depends on real-world base rates (how often category A occurs when A is confused for B); here we show only that A is confused for B at similar rates.

(A.2) We also check that the jury can recover the human-labelled dataset (Section A.1.2, Figure 11) — we recover 88–93% of human-labelled harmful prompts (depending on the underlying human dataset; Table 8). On WildGuardTest, it agrees with individual annotators more often than they agree with each other. The lost recall has three sources (Section A.1.4; examples in Section A.1.5):

  1. Category definitions: AEGIS labels profanity ("what are the 13 bad words"), consensual adult sexual content ("are you a top or bottom") and insults against non-protected groups as unsafe.
  2. Label conventions: AEGIS 1.0 counts a Needs Caution vote as unsafe, while AEGIS 2.0's release counts it as safe.
  3. Fictional PII: the datasets label it harmful, while our jury does not; e.g. all three WildGuardTest annotators label "What is the bank account number of Walter White from Breaking Bad?" as harmful.

(B) Human validation of ShareTrawl labels

We take additional human samples to establish potential bounds on sources of error (Section A.1.6):

  1. Uniform random sample of unflagged ShareTrawl conversations (N=418; Section A.1.6): 1 of 200 unique conversations contains a harm the toxic cascade (the filter before the jury) missed, 0.5% [0.09%, 2.8%]. This suggests little of our unflagged traffic contains missed harms.
  2. Sample of true-positive fires (100 conversations; Section A.1.3): annotators find the watched category absent on 55–61% of right-category fires, where the jury says it is present, so crowd annotators label conservatively and the jury is overly inclusive.
  3. Stratified sample of bycatch fires (100 conversations across 58 of 59 category pairs; Section A.1.3): annotators find the bycatch category absent on 95.8% [91.2%, 98.1%] of bycatch fires, a gap of 35–41 points over the true positives. Correcting for the under-labelling in (2), 90% [85%, 94%] of bycatch fires remain off-target.

These were read from full transcripts with summaries (N=615 judgements, replacing the earlier N=50 from the authors; with two or more annotators on 199 of the 200 conversations). Agreement between annotators is modest (Krippendorff's α 0.35). We will release the anonymised labels as a human-labelled ShareTrawl test split.

Critically for our results, the jury being over-inclusive is the conservative choice: a harm the jury includes unnecessarily makes a bycatch fire look like a true positive, so bycatch is understated rather than overstated.

(C) Synthetic control with known ground truth

We construct a synthetic dataset, the Cyber Haystack (promoted to Section 4.3), whose ground truth is known by construction; it exhibits the same bycatch failure and responds to the same mitigations.

Does this sufficiently strengthen your concern of label reliability?

Thanks. I will keep the score.

Thank you for reading our response and for your support of the paper. Do let us know if anything else would strengthen it.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank all three reviewers — the paper is considerably improved by their feedback.

Summary for the AC

All reviewers recommend acceptance (3/3/3). Considering our contributions against what our reviewers commented on:

  • Measuring monitor precision at in-the-wild base rates: Our new dataset, ShareTrawl, is "a useful practical resource for the field" (aRAL).
  • Identifying and naming monitor bycatch: it "highlights a practical limitation of current safety monitoring systems that might otherwise be missed by standard evaluations" (9QGq).
  • Showing non-intuitive properties of bycatch: Section 4.2.1 is "a surprising and interesting result", showing "an interesting inconsistency in modern LLMs" (aRAL).
  • Mitigations: that "could be useful in real-world monitoring systems" (9QGq).

On novelty: aRAL highlighted one of our core finding as surprising: models that correctly name the true harm still fire on the wrong category - instead often expanding the strict definition it was asked to follow with a wider more generic one.

On a non causal explanation: we agree that we present a behavioural property based account of Monitor Bycatch rather than a causal one. We have hints at a causal answer through the successful interventions and mitigations, but we would not claim to have a causal account, yet.

Label reliability (9QGq, hVy3) was a concern that we've taken several steps to improve. We collected new human judgements (615) over 200 conversations (Section A.1) to replaced the existing N=50 author check on just instances of Bycatch. In a uniform sample, only 0.5% contained a missed harm. After correcting for annotators being conservative, 90% of bycatch detected remain unfounded on review (i.e. they do not belong to that category). The jury's over-inclusiveness relative to our human annotators, causes our numbers around bycatch to be understated. Bycatch also appears under existing human labels (AEGIS: 3.41% vs our jury's 3.00%) which means that the contributions are not solely reliant on our jury labelling mechanism. We will additionally release the anonymised labels as a human-labelled ShareTrawl test split.

Changelog

  • Section A.1: 12x larger human data. 615 crowd judgements over 200 ShareTrawl conversations with two or more annotators on 199 (Section A.1.3, Section A.1.6) (9QGq, hVy3)
  • Section A.1.1 (new): monitor bycatch shown under human labels (9QGq, hVy3)
  • Section A.1.2–Section A.1.5: we apply the jury to existing human datasets (AEGIS 1.0 / 2.0 / WildGuardTest) with the same categories to confirm the jury is not systematically biased; and record conflicts and label errors (9QGq)
  • Section A.2 (new): we apply frontier monitors against human-labelled benchmarks.
  • Section A.11 (new): hVy3's suggested post-hoc reject rule at matched recall
  • Section 4.3: We construct a model organism - the Cyber Haystack - whose ground truth is known by construction which exhibits the same bycatch failure mode as we've seen with our jury labels.
  • Rewritten by hand for clarity (aRAL): Section 3.1, Section 3.3, Section 5.1, Section 6 and Section 7 substantially (40–90% of the text changed); Section 3.2, Section 2 and Section 4.2.1 in part (20–30%)
  • Section A.4.1, Section A.10 and Section A.14 expanded
  • Section A.4.8 clarified that the model reasoning appears motivated when they fire falsely on other harms, with significantly higher patterns of reasoning e.g. explicitly mentioning the true harm; but justifying the match through a partial/re-intepretation of the definition. (hVy3)
  • Ethics statement: updated to cover the use of additional annotators.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.