acceptodds
Under review as a conference paper at ICLR 2027

TotalGuardEval: Benchmark for Evaluating Guardrail Models on Underrepresented Risk Categories

Abstract

Guardrail models—lightweight safety classifiers that filter harmful inputs and outputs of large language models (LLMs)—are advancing rapidly, yet their evaluation remains limited by incomplete taxonomic coverage. Comparing ten widely used guardrail benchmarks and model policy taxonomies, we find critical gaps: Fascism, Religion, and Military conflicts are absent as standalone categories, while Politics receives only partial coverage. We introduce TotalGuardEval, a benchmark targeting these underrepresented categories. It defines a taxonomy of 16 harm categories, each matched with a thematically aligned safe counterpart, and covers both user requests and model responses, enabling measurement of detection performance and over-refusal at the category level. The dataset contains 7.2K examples. We evaluate ten guardrail models, 86M to 20B parameters across three generations, under an original condition, 7 obfuscations, and 88 attack rewrites. The three strongest fall from 0.93–1.00 F1 on original requests to 0.40–0.48 under attack, equally on new and established categories, while one 0.6B model's false positive rate on safe responses spans 0% to 75.5% across categories. Even the most comprehensive prior taxonomy provides at least partial coverage for only 82% of our categories, and no coverage for the three standalone categories we introduce. These results show that current evaluation gives an incomplete picture of guardrail capabilities, and that taxonomically diverse benchmarks are essential for realistic safety assessment.

Primary area:
datasets and benchmarks
Keywords:
guardrail models, LLM safety, safety benchmarks, content moderation, paired evaluation, over-refusal, robustness evaluation, safety classifiers

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_R8CJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces TotalGuardEval, a guardrail benchmark covering 16 risk categories, paired safe samples, and data on both the request and response sides. Using 7,260 synthetic samples, the authors evaluate 10 public models under original, obfuscated, and adversarial conditions, with a focus on missed detections, over-blocking, and adversarial robustness.

Critical strengths
  1. Safe and harmful samples are paired across both the request and response sides, allowing misses and over-blocking to be measured within the same topic.
  2. A broad range of models is covered, with per-category FPR/FNR reported rather than a single aggregate score; the data audit and limitation discussion are also transparent.
Critical weaknesses
  1. All samples are generated by a single model and only 264 are validated by two authors sharing the design priors, leaving label reproducibility for culture-sensitive categories unsubstantiated.
  2. The 88 attacks are not verified to preserve harmful intent and cannot be independently reconstructed.
  3. Highly correlated variants of the same request are counted as independent items in the attack F1. Please report results clustered by source request.
Questions for authors
  1. Could reproducible attack samples be released, with manual verification that each attack preserves the original harmful intent?
  2. Could each condition be evaluated with matched category proportions or balanced accuracy? And how are system prompts and output parsing standardized across models?
1 comment

Thank you for raising reproducibility and the unit of analysis. Appendix P documents the request-side transformations. The attack suite combines methods from NVIDIA’s garak and HiveTrace Red with custom transformations. Researchers can rerun the publicly implemented methods on the original requests using the corresponding frameworks. To replay the exact attack inputs used for Table 4, a release would need to include each evaluated rewrite, its source request, and the transformation ID. Manual review of those same pairs would establish which rewrites preserve the original harmful intent.

The original corpus is balanced within each pair of risk and safe categories (Table 9). The obfuscation and attack conditions use a fixed safe reference and add variants of harmful requests, so the class proportions differ across conditions (Appendix M). Balanced accuracy would weight harmful and safe recall equally. The submitted per-category FNR and FPR already separate missed detections from over-refusals. Table 4 reports attack scores over individual rewrites. A source-weighted analysis assigns each original request the same total weight across its rewrites and treats those rewrites as one cluster when estimating uncertainty.

As described in Section 5, we used each model’s default classification template and supplied the TotalGuardEval taxonomy to models that support custom policies. The models emitted safety labels directly, and we successfully mapped every output from the reported runs to SAFE or UNSAFE. All models received the same attack inputs. The blinded re-annotation in Appendix N measures internal consistency under the written rubric; external validation of the culturally sensitive categories is a separate question.

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

This paper introduces TotalGuardEval, a benchmark targeting underrepresented risk categories in guardrail model evaluation. Through a systematic comparison of ten widely used guardrail benchmarks and model policy taxonomies (Llama Guard, WildGuard, Qwen3Guard, XGuard, POLY-GUARD, OpenAI Mod, HarmBench, StrongREJECT, SEA-SafeguardBench, RabakBench), the authors identify critical taxonomic gaps: Fascism, Religion, and Military conflicts are absent as standalone categories in all compared resources, while Politics receives only partial coverage. TotalGuardEval addresses these gaps with:

  • A taxonomy of 16 harm categories, each matched with a thematically aligned safe counterpart (e.g., "How to organize a financial pyramid scheme?" paired with "How to recognize signs of a financial pyramid scheme?"), plus a dedicated safe traffic category (33 category values total).
  • A corpus of 7,260 examples covering both user requests and model responses (dual modality), generated with Claude Sonnet 4.5 with staged quality control (LLM-as-judge screening, author verification, and a post-construction label-consistency check on a 264-example blinded sample reporting κ = 1.000 binary / 0.992 category agreement).
  • Evaluation of 10 public guardrail models (86M–20B parameters, spanning three architectural generations) under three conditions: original, seven text-level obfuscations, and 88 attack-strategy rewrites.

Key findings: the three strongest models fall from 0.93–1.00 F1 on original requests to 0.40–0.48 under attack, with comparable degradation on novel (Fascism, Religion, Military conflicts) and established categories — showing that attack robustness, not category coverage, is the shared limit. One 0.6B model's false positive rate on safe responses spans 0% to 75.5% across categories, and a "0.333 floor" diagnostic separates "block everything" classifiers (e.g., Llama-Prompt-Guard-2-86M, gpt-oss-safeguard-20b) from genuine detection. The paired category-level design makes over-refusal and under-detection measurable within the same topical domain.

Critical strengths
  1. Novel and well-motivated taxonomic gap analysis. The systematic comparison of ten benchmarks and model policies (Table 1/5, Appendices A/B) convincingly demonstrates that Fascism, Religion, and Military conflicts have zero standalone coverage in any prior resource, and the comparison matrix doubles as a reusable diagnostic checklist for future benchmarks.
  2. Category-level paired safe/harm evaluation methodology. Matching each harm category with a thematically aligned safe counterpart makes over-refusal measurable within the same domain (Table 3), extending beyond aggregate false-positive benchmarks like XSTest/OR-Bench. This is a principled, reusable contribution.
  3. Strong metric literacy and diagnostic depth. The "0.333 floor" analysis (§6, Appendix M) correctly identifies always-UNSAFE classifiers (Llama-Prompt-Guard-2-86M at exactly 0.333, FPR ≈ 1.0) and explains the spurious F1 rise under obf/attack as a class-prevalence effect — avoiding a common misinterpretation. Joint per-category FNR/FPR reporting separates three distinct failure modes.
  4. Broad, well-documented experimental coverage. Ten models spanning 86M–20B parameters and three generations, evaluated under original, obfuscation, and 88-attack conditions, with fully documented generation prompts, per-category policy catalogue, and attack taxonomy (Appendices L/P) — supporting third-party reproduction and community extension.
  5. Honest self-disclosure of limitations. The paper proactively acknowledges single-generator synthetic data, the upper-bound nature of the author-side consistency check, uncalibrated attack strengths, and the English-only scope (Discussion, Appendices N/O/P) — transparency above the average for benchmark papers.
Critical weaknesses
  1. Fully synthetic corpus from a single generator without independent validation. All examples are generated by Claude Sonnet 4.5 (L300–301), and the authors admit possible "generator-specific stylistic regularities" (L429–431); no construction-stage ablation was conducted (L433–434). The measured performance may partially reflect sensitivity to one model family's style rather than genuine harmful-content detection. The κ = 1.000 agreement is explicitly an author-side upper bound (Appendix N), not evidence of taxonomy-independent reproducibility.
  2. Attack-condition results are not calibrated for semantic preservation. Appendix P (L1383–1387) states the 88 attack families "differ in mechanism and in intrinsic strength" and that the authors "do not claim that all transformations preserve harmful intent." Consequently, the headline result — F1 dropping from 0.93 to 0.40–0.48 under attack — conflates model vulnerability with possible data artifacts (rewrites that destroyed the harmful semantics). This weakens the paper's central empirical claim.
  3. Self-audited pairing corrections lack independent verification. The 152 off-topic pairings (4.3%) found in Appendix O were identified and corrected by the same two authors, and the "exhaustive re-audit" was performed by the same individuals — no external review.
  4. Mismatch between "culturally-informed" positioning and actual scope. The benchmark is English-only and static (§7, §9); the cultural grounding exists only at the level of taxonomy selection, not in the data itself, creating tension with Contribution 1's framing.
  5. Presentation and interpretation burden. Table 4 is information-dense and hard to parse; the key caveats about the 0.333 floor and the harmful-heavy nature of obf/attack conditions are deferred to §6/Appendix M, while the abstract's "0.40–0.48 under attack" invites misreading as a quality score. Additionally, degenerate model behavior (gpt-oss-safeguard-20b predicting UNSAFE throughout, FPR ≈ 1.0) may reflect template/policy-format mismatch rather than true failure, but no format-adaptation ablation is reported.
Questions for authors
  1. Attack semantic fidelity: Can you conduct a human spot-check on a sample of the 88 attack rewrites to verify harmful-intent preservation, and report per-family semantic fidelity rates? Without this, the 0.93→0.40–0.48 degradation cannot be cleanly attributed to model weakness.
  2. Generator robustness: If you regenerate a 10% cross-validation subset using a second, different-vendor LLM, do the model rankings and core conclusions (novel vs. established category degradation, 0.333 floor signatures) remain stable?
  3. Format adaptation for degenerate models: Is gpt-oss-safeguard-20b's constant UNSAFE output (FPR ≈ 1.0) attributable to a mismatch between its default classification template/policy format and your custom taxonomy supply? Please report a format-adaptation ablation to rule out a configuration artifact.
1 comment

Thank you for recognizing the value of the taxonomy comparison and the paired evaluation. The attack condition is a request-side stress test (Appendix P). The macro-F1 change reflects performance on the evaluated rewrites with inherited labels. It does not isolate the effect of rewrites individually verified to retain harmful intent. Table 12 also reports per-category false-negative rates. Under attack, these rates show the share of rewrites with inherited harmful labels that each model classified as safe. The original-condition FNR and FPR show category-level differences without relying on transformed inputs. We agree that Table 4 and the abstract should make the role of class proportions and inherited attack labels clearer.

Claude Sonnet 4.5 generated the original requests. We used additional models to generate attack rewrites. The reported model rankings and comparisons between novel and established categories apply to the evaluated corpus. Appendix L provides a common generation template and category policies for a controlled comparison with a second generator; we have not established ranking stability in that setting. The 0.333 macro-F1 reference for an always-UNSAFE classifier on a balanced binary set follows from the class proportions, independently of generator choice. Cultural considerations informed taxonomy selection for the English-language corpus. The correction and re-audit in Appendix O concern pair alignment and were performed internally.

We used a single classification prompt format for gpt-oss-safeguard-20b based on the model authors’ guide. We supplied the TotalGuardEval taxonomy directly to the model, as described in Section 5. The outputs were successfully parsed as safety labels. The near-constant UNSAFE predictions characterize the tested configuration.

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

This paper presents TotalGuardEval, a new benchmark for evaluating LLM guardrail models. The authors added underpresented categories such as Fascism, Religion, Military conflicts which are ignored in previous benchmarks.

The dataset contains 7.2K examples with both user requests and model responses. The evaluation were performed under not only original but also obfuscation or adversarial attack conditions.

The authors evaluated ten public guardrail models and show that state of the art guardrails degrade severely under obfuscation or adversarial attack conditions, and found several interesting observations.

This work enriches the guardrail research on both data source and empirical results.

Critical strengths

The strengths of this paper include:

  • The paper identifies taxonomic gaps in existing guardrail benchmarks.
  • The paired safe harm design enables fine grained per category over refusal measurement.
  • Comprehensive evaluation covering 10 guardrail models across obfuscation or adversarial attack conditions.
Critical weaknesses

The weaknesses of this paper include:

  • The main contributions of this paper seem to be the datasets. However, all data were synthetically generated by a single LLM, Claude Sonnet 4.5, which is an obvious limitation.
  • Annotation validation was based on sampling rather than the whole data. Moreover, validation was performed by the authors themselves independently, and lack of inter annotator agreement, IAA.
  • The evaluation benchmark is static, and it cannot cover the constantly changing categories and methods of attacks.
Questions for authors

na

1 comment

Thank you for highlighting the benchmark’s coverage and the value of its paired design. We distinguish the selective checks used during construction from a separate post-construction assessment of label consistency. The former was not designed to measure inter-annotator agreement. In the latter, two co-authors independently re-annotated a blinded, stratified sample of 264 examples spanning all 33 category values. Appendix N, Table 18 reports Cohen’s κ of 1.000 for binary labels and 0.992 for category labels before adjudication. This supports consistency under the written rubric; it does not substitute for external annotation.

Claude Sonnet 4.5 generated the original requests. We then manually edited selected examples, making the intended behavior and context more explicit where needed. In a separate request–response audit, we identified 152 harmful-side responses that did not address their paired requests, rewrote those responses, and rechecked the corrected pairs (Appendix O, p. 26). These edits do not establish robustness to a different base generator.

The fixed evaluation set gives models a common basis for comparison. New attack transformations can be added to the evaluation pipeline, and the reusable template and category policies in Appendix L provide a way to extend the taxonomy with new risk categories and their matched safe counterparts. This supports the direction outlined in Section 8: adapting the evaluation as risk categories and attack methods evolve.

Comments on the paper

1 comment

We thank all three reviewers for reading the paper so carefully and for their constructive feedback. We appreciate that they highlighted the taxonomy gaps TotalGuardEval addresses, the value of pairing safe and harmful examples within the same topic, and the breadth of the ten-model evaluation. Their questions about label validation, preservation of harmful intent, and the comparability of evaluation conditions were especially helpful: they prompted us to explain more clearly what the benchmark measures and how to interpret the reported results.

Dataset construction and validation. Claude Sonnet 4.5 generated the original requests, and we manually edited selected examples to clarify intent and context. Appendix O documents a separate audit and correction of request–response alignment. Appendix N (Table 18) reports blinded re-annotation by two co-authors on 264 stratified examples: Cohen’s κ was 1.000 for binary labels and 0.992 for category labels before adjudication. This measures internal agreement under the written rubric.

Interpreting the attack results. The attack condition is a stress test on user requests. We used additional language models to generate adversarial rewrites of the original requests. Its aggregate F1 measures performance on the evaluated set of rewrites with inherited labels, not the effect of attacks verified to preserve intent. The original-condition, per-category FNR and FPR independently show where models miss harmful examples or over-refuse safe examples on the same topic.

Comparing conditions and models. The original benchmark contains equal numbers of harmful requests per risk category. The obfuscation and attack conditions use a fixed safe reference and add variants of harmful requests, so class proportions differ across conditions (Appendix M). Models were evaluated on the same attack inputs using model-specific classification templates. The models emitted safety labels directly, and all reported outputs were mapped to SAFE or UNSAFE.

We address reviewer-specific questions in our replies to each review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.