acceptodds
Withdrawn from ICLR 2027

Benchmarking Jailbreak Guardrails for Embodied Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Embodied agents powered by large language models and vision-language models are increasingly deployed in physical environments, but jailbreak attacks can induce these agents to perform physically harmful actions. A growing number of guardrail methods have been proposed to intercept dangerous behavior before it is executed, yet existing safety benchmarks evaluate the embodied models themselves, leaving it unclear how well these guardrails actually defend an embodied agent in practice. We present the first systematic evaluation of jailbreak guardrails for embodied agents. To compare guardrails under identical conditions, we build a pluggable evaluation framework that treats the embodied agent as a fixed backend and each guardrail as a module that can intervene at the perception, planning, or control stage. We subject six representative guardrails to template-based and automated jailbreak attacks as well as safe instructions, and assess them at the system level along three dimensions: defense effectiveness, measured by the bypass rate and the hazard success rate in the simulator; usability, measured by the false-positive rate and the task completion rate on safe instructions; and efficiency, measured by the latency overhead added at runtime. Experiments on guardrails that span different intervention stages, decision mechanisms, and input modalities reveal a clear trade-off among the three dimensions, and show that no single guardrail dominates in all settings. We further analyze how intervention stage, decision mechanism, and input modality shape safety outcomes, and we offer practical guidance for selecting and designing guardrails for embodied agents.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Embodied Agent, Jailbreak Attack, Guardrail

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Me3eOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper compares different jailbreak guardrails of embodied agents with the same backend LLMs (GPT-4o and DeepSeek-V3.2). The paper first defines the benchmark dimensions, including (a) intervention phase; (b) decision mechanism; (c) input modality; and (d) replanning capability. It then runs experiments on SafeAgentBench with 6 guardrail frameworks. Experiment results show several observations and insights: (1) visual information leads to worse guardrail results (Llama Guard 3 and EMBGuard); (2) better defense compromises usability; (3) interception at the planning/control stage (more downstream) induces higher latency; and (4) rule-based has lower latency compared with LLM or agent-based guardrails.

Critical strengths

1. Comprehensive evaluation of the existing jailbreaking guardrail framework.

I appreciate the author's effort in conducting this scale of experiment and also the engineering efforts.

The paper provides a comprehensive evaluation of 6 jailbreaking guardrails with fixed LLM models (GPT-4o and DeepSeek-V3.2). The experiment results are discussed in detail. It's also interesting to see that "adding vision information to the guardrails harms the performance", which is the most interesting finding of this paper.

However, with that being said, I think the paper has critical weaknesses, as discussed below, which make it difficult for me to flip the decision even with this strength.

Critical weaknesses

1. The findings and insights are mostly expected and show limited significance.

The paper discusses several insights (Section 4.3); however, most of them are expected (not surprising or significant to a broader ICLR audience).

  • The defense-usability insight is a well-known trade-off in the area. A guardrail showing better defense capability brings down the clean task performance is already a well-known issue, broadly beyond jailbreaking embodied agents, including adversarial robustness [1] and LLM safety [2], etc. Therefore, this insight adds little new knowledge to this community.

  • Similarly, the insight that "guardrails focused on planning/action have higher latency" is also largely expected: guardrails check every action of the embodied LLM, which naturally requires more compute and thus is slower than guardrails that check at a higher abstraction level. The insight that "Rule-based guardrails are faster than LLM-based guardrails" is even more straightforward.

The finding that "vision harms guardrail performance" is the most interesting one, but I have concerns related to the rigor of this finding (discussed below)

2. The rigor of the claim and insight is questionable, and most of the insights are hypotheses without further evidence.

  • First, the rank in Figure 2 takes the average of different metrics (e.g., bypass rate, false positive, etc.). It looks problematic to me to simply average those metrics numerically because they are computed in different ways, with different scales (even with normalization), and some have strong correlations (e.g., bypass rate and hazard success rate). Therefore, any insight drawn from this ranking is also questionable.

  • Second, the paper claims "vision harm guardrails": EMBGuard (T + V) < EMBGuard (T) and Llama Guard (T + V) < Llama Guard (T) based on the above rank. Since the rank itself is not rigorous, so do the conclusion. For example, EMBGuard (T + V) shows a much lower false positive rate and a higher completion rate than EMBGuard (T). How the ranking is averaged largely influences how the results are interpreted. Also, Llama Guard is not designed for embodied agent tasks; thus, it's unknown how solid the conclusion drawn from it is.

  • Additionally, the insight section has a lot of hypotheses (e.g., "we attribute xxx to yyy", "one possible explanation...") from the main results without further analysis. A concrete example is at Line-401: The paper claims "including vision dilutes attack signal from instruction" without showing further evidence, such as guard model attention scores; at Line-406, the paper claims "embodied guardrails inherit this fragility", a claim largely beyond what the results can support.


Therefore, I think the paper is not publication-ready for ICLR and can be further improved by (1) providing more interesting insight discussion by digging into the results more and (2) revising or justifying why the current rank is scientifically rigorous in Figure 2.

Reference

[1] (ICLR'19) Robustness May Be at Odds with Accuracy

[2] (ACL'24) The Art of Defending: A Systematic Evaluation and Analysis of LLM Defense Strategies on Safety and Over-Defensiveness

Questions for authors

Please see the suggestion above.

Reviewer_F6PROpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper compares external guardrails for protecting embodied agents against jailbreak attacks. The authors organize existing methods along four dimensions: intervention phase, decision mechanism, input modality, and replanning capability. They introduce a modular evaluation framework and test six guardrail methods against five jailbreak attacks, together with benign instructions, using GPT-4o and DeepSeek-V3.2 as the agent backends. Experiments are conducted on SafeAgentBench tasks in AI2-THOR, measuring defense effectiveness, benign-task usability, and runtime cost. The reported results show trade-offs across these criteria, with no single method performing best on all of them. The paper also reports weaker defense when visual inputs are added in the tested configurations and discusses possible explanations and implications for guardrail design.

Critical strengths

The paper addresses a practical problem in embodied AI safety. I appreciate the effort to compare several guardrail methods in a common evaluation setting and to consider benign-task performance and runtime cost alongside defense effectiveness. These are useful considerations for anyone choosing a guardrail for deployment. The distinction between bypass rate and hazard success rate is also worth studying, as it draws attention to what happens during execution and whether harmful outcomes actually occur.

Critical weaknesses
  1. The novelty of the study is not sufficiently established, and the discussion of closely related work is limited. For example, RoboJailBench, which is cited in the manuscript, already provides a unified evaluation of attacks and defenses, measures both safety and benign-task utility, and includes RoboGuard. However, the manuscript does not adequately discuss how its evaluation differs from such closely related efforts before presenting this area as an unresolved gap. The authors should clarify the motivation for the study, identify the specific limitations of existing benchmarks, and explain what this work adds.

  2. The experimental evaluation does not sufficiently support the breadth of the paper’s claims. The proposed taxonomy covers four dimensions: intervention phase, decision mechanism, input modality, and replanning capability. I would like to see more detailed experimental designs and analyses along each of these dimensions, with a clearer explanation of how the results support the corresponding conclusions.

  3. Several details needed to assess the reliability of the results are missing. The main reported rates should be accompanied by effective sample sizes, the number of independent runs, and confidence intervals. The authors should also report how many tasks are evaluated by the rule-based evaluator and by the GPT-4o judge, respectively, and explain how the reliability and consistency of these judgments are assessed. These details could be included in the appendix to improve the transparency and credibility of the evaluation.

  4. I found the relationship between the definition of bypass and the interpretation of replanning results confusing. On page 4, bypass is defined as an episode in which the guardrail never intervenes. On page 6, however, the authors interpret RoboGuard’s bypass rate of 82.05% and hazard success rate of 23.85% as evidence that replanning helps prevent harm. Under the stated definition, an episode involving guardrail-triggered replanning should already count as an intervention. Moreover, a bypass that does not result in harm could also reflect planning or execution failure. The authors should reconcile the definition with this interpretation, clarify how replanning episodes are counted, and explain what evidence supports attributing the reduction in harm to replanning.

  5. I think the explanation that visual inputs dilute the safety signal deserves some further experimental discussion. It would be helpful to include additional experiments, together with a few visual examples or visualizations, to make it easier to understand what is happening and when this effect occurs.

Questions for authors
  1. What specific gap does this benchmark address beyond closely related work such as RoboJailBench? I would appreciate a clearer explanation of which contributions go beyond evaluating more guardrail methods.

  2. How are episodes involving replanning counted in the bypass rate? Given the definition on page 4, I am unsure how to interpret RoboGuard’s 82.05% bypass rate and 23.85% hazard success rate as evidence for the benefit of replanning. What evidence separates this benefit from cases where planning or execution simply fails to produce harm?

  3. Which conclusions about intervention phase, decision mechanism, input modality, and replanning capability are supported by comparisons that isolate the factor being studied? In particular, what evidence supports the proposed explanation that visual inputs dilute the safety signal?

  4. How stable are the main results across repeated runs, and what are the effective sample sizes and confidence intervals? Also, how many tasks are assessed by each evaluator, and what checks have been made on the reliability and consistency of the judgments?

Reviewer_1zGROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a systematic evaluation framework for jailbreak guardrails in embodied agents. The authors categorize existing guardrails along four dimensions: intervention stage, decision mechanism, input modality, and replanning capability. They evaluate six representative guardrails against five jailbreak attacks using GPT-4o and DeepSeek-V3.2 as planning models in AI2-THOR.

The evaluation considers defense effectiveness (bypass rate and hazard success rate), usability (false-positive rate and task completion rate), and efficiency (runtime latency). The results reveal trade-offs among these dimensions. The authors also find that adding visual inputs reduces average defense effectiveness for the evaluated model-based guardrails across both planner backends, and analyze how intervention timing and decision mechanisms relate to runtime costs.

Critical strengths
  1. System-level evaluation of guardrails. Rather than relying solely on detection rates, the paper evaluates downstream task outcomes in simulation. For tasks with goal-state annotations, outcomes are assessed using final simulator states; otherwise, an LLM-based judge compares executed action trajectories against reference plans. Separating bypass rate from hazard success rate provides useful information about detection performance and downstream safety outcomes.

  2. Multi-dimensional evaluation under a common framework. The paper compares guardrails with different architectures and intervention stages using a fixed embodied-agent backend. Considering defense effectiveness, usability, and efficiency together reveals practical trade-offs that would be overlooked by evaluating safety alone. Evaluating two planner backends also helps assess the consistency of the observed findings.

Critical weaknesses
  1. The evaluated attacks are not adaptive to the guardrails. All five attacks target the embodied agent rather than the guarded system itself. Although this provides a useful comparison against existing jailbreak attacks, it remains unclear whether the conclusions hold when an attacker can observe guardrail decisions and adaptively optimize for bypass. This limits the assessment of guardrail robustness against an informed adversary.

  2. The reliability and interpretation of some evaluation outcomes need further clarification. Section 4.1 states that tasks without annotated goal states are evaluated using a GPT-4o-based judge, but the paper does not report how many tasks rely on this judge or provide an independent validation of its judgments. In addition, Section 4.2 attributes RoboGuard's relatively low hazard success rate despite its high bypass rate to replanning. However, the difference between these two metrics alone does not establish the effectiveness of replanning, since harmful outcomes may fail to occur even without intervention. Clarifying the paired evaluation protocol and reporting outcomes before and after replanning would strengthen these conclusions.

Minor suggestions:

  1. The effects of individual guardrail design factors are not fully isolated. Different guardrails often differ simultaneously in intervention stage, decision mechanism, and replanning capability. Therefore, comparisons across different methods do not necessarily establish the individual contribution of each design factor. The paper could more clearly distinguish controlled comparisons from observational findings.

  2. The overall leaderboard depends on metric weighting and normalization. Figure 2 combines five metrics into a single ranking, but their relative importance may vary across deployment scenarios. A sensitivity analysis with different metric weights or a Pareto-front comparison would make the ranking more informative.

  3. Appendix A describes BadRobot-CJ as an iterative attack, which is inconsistent with its template-based description in Section 4.1.

Questions for authors
  1. Have the authors considered guardrail-aware adaptive attacks that optimize malicious instructions using feedback from the guardrail itself? If not, how should the robustness conclusions be interpreted under this stronger threat model?

  2. What proportion of the evaluated tasks relies on the GPT-4o-based judge rather than the rule-based simulator evaluator? Has the judge's reliability been validated against independent annotations or other evaluation methods?

  3. Regarding RoboGuard, how are bypass, intervention, and replanning outcomes counted? Under GPT-4o, the reported bypass rate is 82.05%, while hazard success decreases from 47.74% without a guardrail to 23.85% with RoboGuard. Are the guarded and unguarded evaluations paired, and how much of this reduction can be attributed specifically to replanning?

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

Related papers

Discussion (0)

Sign in to comment.