acceptodds
Under review as a conference paper at ICLR 2027

SafeCoEvo: Co-Evolving Safety Harnesses and Guards for LLM Agents at Test Time

12 views

Abstract

LLM agents deployed in real-world environments continually encounter new tasks and safety risks, while execution feedback typically becomes available only after each task is completed. However, existing self-evolving approaches commonly rely on multiple rounds of optimization over fixed and repeatedly accessible task distributions, fundamentally differing from test-time adaptation in real-world deployment, where only experience accumulated from past tasks can be used to improve safety decisions on future unseen tasks. To address this limitation, we propose **SafeCoEvo**, a test-time Harness–Guard co-evolution framework for LLM agent safety that enables the external safety system to continually adapt from accumulated runtime experience. SafeCoEvo jointly improves two complementary safety capabilities at different timescales: **S-Harness** rapidly externalizes recent runtime experience into updatable explicit safety knowledge that can promptly influence subsequent tasks, while **GuardVPO** internalizes accumulated runtime safety experience over a longer timescale into parametric risk-judgment capabilities. By combining short-term rapid adaptation with long-term capability consolidation, SafeCoEvo continually improves the agent's safety capabilities, reducing the unsafe outcome rate by **10.05%** while improving the task success rate by **12.15%** over the strongest baseline, thereby achieving simultaneous gains in safety and task utility.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Agents, Agent Safety, Agent Security, Test-Time Evolution, Safety Harness

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_5FqxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Existing agent-safety methods typically rely on static guards or harnesses, or on self-evolution with repeatedly accessible fixed datasets, whereas deployment more closely resembles continual test-time adaptation in which tasks arrive sequentially and only past trajectories and feedback are available. The authors address this setting with a dual-timescale Harness–Guard co-evolution framework: S-Harness updates prompts, validated memory, safety skills, permission policies and guard rules after each task, while GuardVPO periodically internalizes accumulated experience into the Guard through event-level safety feedback, asymmetric rewards, PPO-style clipping and KL regularization, with the task-execution model kept frozen.

Critical strengths
  • Interesting dual-timescale design. GuardVPO addresses an important dimension of agent safety, while SafeCoEvo cleanly separates short-term adaptation through S-Harness from long-term parametric consolidation through GuardVPO.
  • Practical continual test-time formulation. The paper studies a realistic deployment setting in which tasks arrive sequentially and adaptation can use only trajectories and feedback from previously observed tasks.
  • Strong generalization. The method maintains strong performance on the held-out test set and yields consistent improvements when either the task-performing model or the Guard backbone is replaced.
Critical weaknesses
  • Limited evaluation on frontier models. The evaluation is restricted to relatively small or mid-scale models. It would be valuable to assess whether SafeCoEvo remains effective on stronger frontier models, such as GPT-5.5 and Opus 5.5.
  • Limited baseline coverage. The comparison set is relatively narrow. The authors should include one or two additional competitive agent-safety baselines.
  • Missing AgentDojo comparison. Given its relevance to agent safety and prompt-injection robustness, AgentDojo should be included as either an evaluation benchmark or comparison point.
  • Notation and presentation. A dedicated notation table would improve readability and make the framework easier to follow, particularly given the interaction between S-Harness, GuardVPO, memory, and event-level feedback.
  • Statistical robustness. The paper should report standard deviations or confidence intervals across multiple runs to better establish the reproducibility and robustness of the reported gains.
  • Reliance on reliable risk feedback. GuardVPO depends on event-level risk feedback, an assumption that may be difficult to satisfy in realistic deployments where safety feedback can be delayed, incomplete, noisy, or entirely unavailable.
  • Long-term memory stability. Because S-Harness continually accumulates explicit safety knowledge, redundancy and conflicts may increase over time. The current evaluation provides limited analysis of memory growth, rule conflicts, catastrophic adaptation, or performance degradation over substantially longer task streams.
Questions for authors
  • Table formatting. Tables should follow standard LaTeX conventions, using \toprule and \bottomrule for cleaner presentation.
  • Equation formatting. Unnumbered equations that are not referenced in the text should use \begin{align*}...\end{align*} rather than numbered environments.
  • Lack of reproducibility and ethical considerations
11 comments
AuthorsOpenReview ↗

Response to Weakness 1: Effectiveness with a Frontier Task-Performing Model

Thank you for this valuable suggestion. To further evaluate the effectiveness of SafeCoEvo on frontier models, we conduct additional experiments using GPT-6.1-terra as the task-performing model, comparing SafeCoEvo, SafeCoEvo (Static), and SHE over 512 episodes.

As shown in Table A, compared with SafeCoEvo (Static), online evolution reduces UOR by 9.18 percentage points while increasing TSR and SUCR by 0.39 and 0.20 percentage points, respectively. Compared with SHE, SafeCoEvo reduces UOR by 14.65 percentage points and increases TSR and SUCR by 6.25 and 4.88 percentage points, respectively. Together with the results for DeepSeek-v4.1-Flash and Qwen3.7-Flash reported in the main text, these findings show that SafeCoEvo’s benefits extend across different task-performing models and that its external safety adaptation mechanism does not depend on a particular backbone. We will include these additional results and experimental details in the revised manuscript and appendix.

Table A. Performance comparison with GPT-6.1-terra as the task-performing model.

MethodUOR ↓TSR ↑SUCR ↑
SHE26.17%29.10%28.52%
SafeCoEvo (Static)20.70%34.96%33.20%
SafeCoEvo11.52%35.35%33.40%
AuthorsOpenReview ↗

Response to Weaknesses 2 and 3: Additional Baseline Comparisons and Cross-Benchmark Evaluation

Thank you for these valuable suggestions. To broaden the baseline coverage, we introduce ShieldAgent as an additional Guard baseline and evaluate the contributions of S-Harness evolution and GuardVPO with this Guard, using Qwen3.7-Flash as the task-performing model. We also expand the external evaluation on AgentDojo and add results on AgentDyn.

Additional Guard baseline. Table A compares SHE, our static configuration with ShieldAgent, and online S-Harness evolution with ShieldAgent. “+ ShieldAgent” denotes replacing the original AgentDoG 1.5 Guard with ShieldAgent. In the two ShieldAgent configurations, the Guard parameters remain fixed, and GuardVPO is disabled. Compared with the static configuration, S-Harness evolution reduces UOR by 9.70 percentage points and increases TSR and SUCR by 12.08 and 11.68 percentage points, respectively. Compared with SHE, it reduces UOR by 13.15 percentage points and increases TSR and SUCR by 18.17 and 13.42 percentage points, respectively. These results supplement the baseline comparisons and support the effectiveness of S-Harness evolution with an additional Guard model.

Table A. Performance on the continual evolution stream using Qwen3.7-Flash as the task-performing model.

MethodUOR ↓TSR ↑SUCR ↑
SHE39.88%22.82%20.83%
SafeCoEvo (Static) + ShieldAgent36.44%28.91%22.57%
SafeCoEvo w/o GuardVPO + ShieldAgent26.73%40.99%34.26%

We further evaluate the contributions of S-Harness evolution and GuardVPO with ShieldAgent on the held-out test set. As shown in Table B, replacing the initial S-Harness with its evolved version reduces UOR by 10.67 percentage points and increases TSR and SUCR by 12.33 and 10.67 percentage points, respectively. With the evolved S-Harness held fixed for this controlled comparison, updating ShieldAgent through GuardVPO further reduces UOR by 5.00 percentage points and increases TSR and SUCR by 3.67 and 5.00 percentage points, respectively. These results support GuardVPO’s additional contribution beyond S-Harness evolution and show that its benefits are not limited to the original Guard model.

Table B. Held-out performance with ShieldAgent before and after S-Harness evolution and GuardVPO.

SettingUOR ↓TSR ↑SUCR ↑
H0+G0H_0+G_032.33%22.67%16.33%
H512+G0H_{512}+G_021.67%35.00%27.00%
H512+G1H_{512}+G_116.67%38.67%32.00%

Here, H0H_0 and H512H_{512} denote the initial S-Harness and the version obtained after 512 episodes, respectively. G0G_0 and G1G_1 denote ShieldAgent before and after the GuardVPO update.

AgentDojo and AgentDyn evaluation. Table 11 in Appendix C of the original manuscript already reports results on a subset of AgentDojo. We expand this evaluation to 249 cases across the Slack and Banking scenarios and additionally evaluate 200 cases from the DailyLife scenario in AgentDyn. In both benchmarks, cases are constructed by combining selected user tasks with injection tasks. We use Qwen3.7-Flash as the task-performing model and AgentDoG 1.5 as the Guard backbone. SafeCoEvo is evaluated after evolution on the original adaptation stream, with both S-Harness and the Guard parameters frozen throughout the cross-benchmark evaluation. No interaction trajectories or feedback from these external benchmarks are used for updates. The results are presented in Table C.

Table C. Cross-benchmark evaluation on AgentDojo and AgentDyn with Qwen3.7-Flash.

BenchmarkConfigurationUOR ↓TSR ↑SUCR ↑
AgentDojoQwen3.7-Flash45.38%75.50%43.37%
AgentDojoSHE6.83%75.10%69.48%
AgentDojoSafeCoEvo0.00%75.50%75.50%
AgentDynQwen3.7-Flash24.00%74.00%56.50%
AgentDynSHE4.00%77.00%73.50%
AgentDynSafeCoEvo0.00%76.00%76.00%

As shown in Table C, no unsafe outcomes are observed among SafeCoEvo’s valid outputs on either benchmark. Compared with SHE, SafeCoEvo reduces UOR by 6.83 percentage points on AgentDojo and 4.00 percentage points on AgentDyn. On AgentDojo, TSR increases by 0.40 percentage points and SUCR increases by 6.02 percentage points. On AgentDyn, TSR decreases by 1.00 percentage point, while SUCR increases by 2.50 percentage points. These results support improved safe task completion in the evaluated external prompt-injection scenarios, with task success rates close to those of SHE.

We will include these additional comparisons, results, and experimental protocols in the revised manuscript and appendix.

AuthorsOpenReview ↗

Response to Weakness 4 and Presentation Comments: Notation, Formatting, Reproducibility, and Ethics

Thank you for your suggestions regarding notation, formatting, reproducibility, and ethical considerations. We have made the following changes in the revised manuscript.

  1. Notation. We have added Table 1, “Notation for SafeCoEvo,” to Section 3.1 (Overview). It summarizes the notation for tasks, execution trajectories, episode-level feedback, accumulated runtime experience, S-Harness states, Guard states, and event-level safety experience. The table also clarifies that Validated Memory is part of the S-Harness state and lists the event inputs, verdicts, and rewards used by GuardVPO. This provides a reference for the task-by-task updates in Section 3.2 and the event-level optimization in Section 3.3.

  2. Table formatting. We have standardized the experimental results tables in Section 4 and Appendices A–D using the three-rule style provided by the LaTeX booktabs package. Specifically, we use \toprule, \midrule, and \bottomrule at the top of each table, below the header, and at the bottom, respectively. Metric groups within multilevel headers are marked using \cmidrule.

  3. Equation formatting. Following the reviewer’s suggestion, we have changed the displayed equations in Sections 2–3 that require no numbering and are not referenced in the text to use \begin{align*}...\end{align*}.

  4. Reproducibility and ethical considerations. We have added a Reproducibility Statement and an Ethics Statement after the AI Use Statement and before the references.

The Reproducibility Statement points to the S-Harness update procedure and GuardVPO optimization objective in Sections 3.2–3.3, the data splits and evaluation protocol in Section 4.1, the GuardVPO optimization curves and validation results in Appendix I, and the Target Agent and Evolution Agent prompt templates in Appendix L.

The Ethics Statement explains that the experiments are conducted in benchmark environments and that SafeCoEvo is intended to provide runtime safety protection for tool-using language model agents. It also discusses the uses of prompt-injection examples and clarifies that the experimental results are reported and interpreted within the scope of the benchmark tasks used.

Reviewer_Vu6MOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies continual test-time safety adaptation for LLM agents, a setting where tasks arrive sequentially over a deployment stream and only the experience accumulated from previously completed tasks can be used to improve safety on future, unseen tasks. The authors propose SafeCoEvo, which keeps the task-performing model frozen and instead co-evolves two external safety components at different timescales: (1) S-Harness, which after every episode uses an Evolution Agent to distill the trajectory and official feedback into five updatable explicit safety components and (2) GuardVPO, which periodically internalizes accumulated event-level safety experience into the Guard's parametric risk-judgment capability via a clipped policy-optimization objective with an asymmetric reward and KL regularization to the initial Guard. Experiments show that SafeCoEvo reduces the unsafe outcome rate by 10.05% while improving the task success rate by 12.15% over the strongest evolving baseline.

Critical strengths

Clear problem formulation. The distinction between existing self-evolving approaches and continual test-time adaptation is clearly articulated and formally defined in Eqs. (1)–(2).

Sensible dual-timescale design. Separating fast explicit adaptation (S-Harness, updated per episode without any training) from slow parametric internalization (GuardVPO, applied only to a small Guard model) is a practical decomposition: it avoids the cost of repeatedly training and redeploying the task-performing model, which is a convincing argument given the increasing scale of agent backbones.

The safety–utility trade-off is explicitly addressed rather than ignored. The asymmetric reward in GuardVPO penalizes missed unsafe events most heavily while still penalizing over-conservative verdicts, and the evaluation reports TSR/SUCR alongside UOR, so the reported safety gains cannot be achieved simply by blocking more actions. The observed simultaneous improvements in both safety and task success support this claim.

Critical weaknesses

The motivation of this paper is unclear. The method appears to depend on having reliable safety feedback after every sample (coming from the evaluator of the dataset samples), yet the authors themselves claim that in real-world deployment feedback is delayed and scarce. In the real world, however, there is no ground truth or safety labels.

Effectiveness of the method. The training uses Agent-SafetyBench and Agent Security Bench, and so does the testing, but these data are possibly from the same distribution. This is especially true for Agent Security Bench: this benchmark is constructed by permutation and combination of injections, so in practice it is very easy for the method in this paper to overfit it. The authors should add some OOD benchmarks, such as AgentHarm, AgentDojo, or AgentDyn, to verify that the method is actually effective.

Missing baselines. For Table 1, if what is being evolved is the guard model, then stronger guard baselines or general model baselines, such as ShieldAgent, should be compared. Table 5 only compares against the AgentDoG 1.5 baseline.

Missing ablations. It is still unclear what role each component of the S-Harness actually plays, and ablation experiments are needed to investigate the contribution of each component.

Questions for authors

See weakness.

8 comments
AuthorsOpenReview ↗

Response to Question 1: Motivation and the Availability of Safety Feedback

Thank you for raising this important concern. We acknowledge that our description of the feedback signals was insufficiently clear and apologize for the confusion. We would like to clarify that this work treats post-hoc feedback as an external input and focuses on using historical interaction experience to improve safety on subsequent unseen tasks. The mechanism for acquiring feedback is not the primary research objective of this work. Our focus is on enabling the agent’s external safety system to evolve continually at test time using experience and feedback accumulated during operation. In our formulation, ftf_t denotes the feedback signal associated with a completed episode, rather than necessarily a ground-truth safety label. The central question is how to translate accumulated interaction experience into improved safety capabilities for subsequent unseen tasks.

The feedback used in our framework does not come exclusively from ground-truth annotations. Depending on the evaluation setting, feedback may also come from LLM-as-a-Judge assessments. Specifically, some of our evaluation signals are obtained through judgments made by large language models rather than ground-truth labels. These signals serve as available supervision for identifying safety-related outcomes and guiding subsequent adaptation; they do not imply that perfect safety annotations are always available in real-world environments.

Prior studies have also investigated how interaction experience can drive continual improvement given available feedback. In agent safety, Membrane [1] uses interaction-level verdicts to update contrastive safety memory, while SafeEvolve [2] uses safety and utility scores from rule-based verifiers to guide harness and policy updates. In broader research on agent adaptation and prompt optimization, ACE [3] updates contextual playbooks through reflection and curation in settings with task feedback, whereas GEPA [4] combines execution trajectories with evaluation feedback to propose and select prompt updates. These studies share a common research perspective: treating available feedback as an input and investigating how experience and feedback can be translated into improved performance on subsequent tasks.

In the revised manuscript, we will further clarify the distinction between acquiring feedback and performing feedback-driven test-time evolution, explicitly describe the feedback sources used in our experiments, and discuss the limitations arising from feedback availability in real-world deployment.

[1] Choi, Minseok, et al. “Membrane: A Self-Evolving Contrastive Safety Memory for LLM Agent Defense.” EMNLP, 2026.

[2] Mao, Qinghua, et al. “SafeEvolve: Harness-Policy Co-Evolution from Agent Experience for Safety Alignment.” arXiv preprint arXiv:2609.02786, 2026.

[3] Zhang, Qizheng, et al. “Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models.” ICLR, 2026.

[4] Agrawal, Lakshya A., et al. “GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning.” ICLR, 2026.

AuthorsOpenReview ↗

Response to Question 2: Effectiveness and Generalization Beyond the Adaptation Benchmarks

Thank you for highlighting the potential risk of overfitting when adaptation and evaluation use the same source benchmarks. We agree that evaluation solely on Agent-SafetyBench and Agent Security Bench is insufficient to fully assess generalization to external scenarios. Under our evaluation settings, attacks in AgentDojo and AgentDyn are relatively easy for existing defenses to block. We therefore selected Agent-SafetyBench and Agent Security Bench as our main benchmarks to evaluate more challenging safety scenarios.

Table 11 in Appendix C of the original manuscript already reports an evaluation on a challenging subset of AgentDojo. To broaden the evaluation coverage, we conduct additional cross-benchmark experiments on AgentDojo and AgentDyn using Qwen3.7-Flash as the task-performing model, comparing SafeCoEvo with the base model and SHE. We evaluate SafeCoEvo after evolution on the original adaptation stream, keeping both S-Harness and the Guard parameters frozen throughout the cross-benchmark evaluation. No interaction trajectories or feedback from the external benchmarks are used for updates. The AgentDojo evaluation comprises 249 cases across the Slack and Banking scenarios, while the AgentDyn evaluation comprises 200 cases from the DailyLife scenario. For both benchmarks, cases are constructed by combining selected user tasks with injection tasks. The results are presented in Table A.

Table A. Cross-benchmark evaluation on AgentDojo and AgentDyn with Qwen3.7-Flash.

BenchmarkConfigurationUOR ↓TSR ↑SUCR ↑
AgentDojoQwen3.7-Flash45.38%75.50%43.37%
AgentDojoSHE6.83%75.10%69.48%
AgentDojoSafeCoEvo0.00%75.50%75.50%
AgentDynQwen3.7-Flash24.00%74.00%56.50%
AgentDynSHE4.00%77.00%73.50%
AgentDynSafeCoEvo0.00%76.00%76.00%

As shown in Table A, no unsafe outcomes are observed among SafeCoEvo’s valid outputs on either benchmark. On AgentDojo, SafeCoEvo achieves a TSR comparable to SHE while improving SUCR by 6.02 percentage points. On AgentDyn, TSR decreases slightly by 1.00 percentage point, while SUCR increases by 2.50 percentage points. Compared with the base model, SafeCoEvo substantially reduces observed unsafe outcomes on both benchmarks while maintaining or improving task success rates.

These cross-benchmark results show that SafeCoEvo improves safe task completion in prompt-injection scenarios beyond the main evaluation benchmarks, providing further evidence of the generalization of its safety capabilities across environments. We will include these additional results in Appendix C of the revised manuscript.

AuthorsOpenReview ↗

Response to Question 3: Additional Guard Baselines

Thank you for this valuable suggestion. We agree that broader baseline comparisons are important for evaluating the effectiveness of SafeCoEvo. Using Qwen3.7-Flash as the task-performing model, we introduce ShieldAgent as an additional Guard baseline and conduct supplementary experiments on the continual evolution stream, the held-out test set, and external Guard evaluation benchmarks.

We first evaluate the continual adaptation of S-Harness while keeping ShieldAgent fixed. As shown in Table A, compared with the static configuration, online S-Harness evolution reduces UOR by 9.70 percentage points and increases TSR and SUCR by 12.08 and 11.68 percentage points, respectively. These results show that the benefits of S-Harness evolution extend to a different Guard backbone.

Table A. Performance on the continual evolution stream with ShieldAgent.

MethodS-HarnessUOR ↓TSR ↑SUCR ↑
SHE (Qwen3.7-Flash)Online evolution39.88%22.82%20.83%
SafeCoEvo (Static) + ShieldAgentFixed H0H_036.44%28.91%22.57%
SafeCoEvo w/o GuardVPO + ShieldAgentOnline evolution26.73%40.99%34.26%

We further examine the contributions of S-Harness and GuardVPO on the held-out test set. As shown in Table B, S-Harness evolution reduces UOR by 10.67 percentage points and increases TSR and SUCR by 12.33 and 10.67 percentage points, respectively. More importantly, with S-Harness held fixed, subsequently applying GuardVPO further reduces UOR by 5.00 percentage points and increases TSR and SUCR by 3.67 and 5.00 percentage points, respectively. This controlled comparison demonstrates GuardVPO’s additional contribution beyond the evolution of explicit knowledge in S-Harness and shows that this benefit is not restricted to a particular Guard backbone.

Table B. Held-out performance with ShieldAgent.

SettingUOR ↓TSR ↑SUCR ↑
H0+G0H_0+G_032.33%22.67%16.33%
H512+G0H_{512}+G_021.67%35.00%27.00%
H512+G1H_{512}+G_116.67%38.67%32.00%

Finally, we evaluate ShieldAgent’s risk assessment performance before and after GuardVPO on R-Judge and ATBench. As shown in Table C, Unsafe F1 and Recall improve by 0.17 and 3.39 percentage points on R-Judge, and by 6.59 and 10.06 percentage points on ATBench, respectively. These results support improved unsafe-event detection on external benchmarks, while lower Safe Specificity reveals a trade-off between unsafe-event sensitivity and safe-class specificity.

Table C. Risk assessment on external Guard benchmarks before and after GuardVPO.

GuardR-Judge Unsafe F1 ↑Recall ↑SafeSpec ↑ATBench Unsafe F1 ↑Recall ↑SafeSpec ↑
AgentDoG 1.585.77%84.40%86.38%79.33%72.75%89.58%
AgentDoG 1.5 + GuardVPO86.23%85.46%85.99%80.61%75.51%88.35%
ShieldAgent83.39%89.56%71.91%51.90%41.25%82.50%
ShieldAgent + GuardVPO83.56%92.95%67.16%58.49%51.31%76.14%

We will include these comparisons and detailed experimental results in the experimental section and appendix of the revised manuscript. These experiments broaden the baseline coverage and further demonstrate that SafeCoEvo’s benefits are not limited to the original AgentDoG 1.5 Guard configuration.

Reviewer_AcnGOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a test-time external safety mechanism for LLM agent safety that continually adapts from runtime experience. It combines two approaches, S-Harness and Guard, each with its own adaptation scheme operating at different timescales. S-Harness provides explicit safety knowledge and behavioral constraints and is updated after each task completion. The experience is abstracted into safety knowledge to be used across future tasks. Guard assesses the risks of events and is updated at a longer timescale. The long-term safety experience is internalized into the Guard's parametric risk-judgment capabilities. The proposed framework was evaluated on benchmarks against several baselines, demonstrating its ability to generalize runtime experience into versatile safety knowledge through test-time adaptation, consequently improving safety and task utility.

Critical strengths
  1. Extensive experiments that demonstrate the versatility of the proposed framework by evaluating it on several benchmarks across different system backbones

  2. Superior performance in terms of safety and task utility compared with other methods

Critical weaknesses

1. Weak necessity of GuardVPO in the proposed co-evolution framework

The experimental results do not sufficiently support the necessity of GuardVPO. In the experiments, GuardVPO is applied only once or twice, making it difficult to assess its effect on performance. In addition, the generalization results show only marginal improvements, which weakens the justification for incorporating GuardVPO.

Questions for authors
  1. Could you explain why the update interval is set to 512 episodes out of a total 1024 episodes? With such a limited number of updates, it is difficult to observe performance trends as the Guard continually adapts. It would be beneficial to evaluate the framework on longer task streams with the same update interval or to consider shorter update intervals (eg. 128, 256) to better demonstrate the effectiveness of continual Guard adaptation.

  2. Could you analyze the performance variation when S-Harness update and GuardVPO are applied simultaneously. In the current experiments, they are updated separately; specifically, only S-Harness is updated during the first 512 episodes, while GuardVPO is subsequently applied with S-Harness fixed. An evaluation with both components continually updated would help clarify the effectiveness of the proposed co-evolution framework.

6 comments
AuthorsOpenReview ↗

Response to Weakness 1: Evidence for GuardVPO’s Contribution

Thank you for raising this concern. We acknowledge that the original 512-episode interval allows few GuardVPO updates within the 1,024-episode stream. We therefore clarify its complementary role and provide additional evidence on repeated updates and generalization.

S-Harness converts runtime experience into reusable explicit safety knowledge and constraints, while GuardVPO incorporates event-level safety experience into the Guard’s parameters. Together, they enable rapid adaptation through explicit knowledge and continual improvement of risk assessment through parameter learning.

Using Qwen3.7-Flash as the task agent, we perform GuardVPO updates every 256 or 128 episodes and evaluate the resulting checkpoints on the held-out test set. For each paired evaluation before and after a GuardVPO update, we use the same S-Harness checkpoint to isolate GuardVPO's incremental contribution. HtH_t denotes S-Harness after tt episodes; GkG_k denotes the Guard after kk updates, with G0G_0 the initial Guard.

With a 256-episode interval (Table A), the first update at fixed H256H_{256} reduces UOR by 9.33 percentage points and increases TSR/SUCR by 6.00/6.33 points. The second update at fixed H512H_{512} reduces UOR by 1.67 points and increases TSR/SUCR by 7.00/8.33 points, demonstrating additional safety and utility benefits at both stages.

Table A. Held-out performance with a 256-episode GuardVPO update interval.

SettingUOR ↓TSR ↑SUCR ↑
H256+G0H_{256}+G_022.00%42.00%36.67%
H256+G1H_{256}+G_112.67%48.00%43.00%
H512+G1H_{512}+G_114.00%48.00%42.00%
H512+G2H_{512}+G_212.33%55.00%50.33%

With a 128-episode interval (Table B), GuardVPO undergoes four updates over 512 episodes. Despite fluctuations in task utility during co-evolution, the fourth update, evaluated with H512H_{512} fixed, reduces UOR by 1.67 percentage points and improves both TSR and SUCR by 6.00 percentage points, providing further evidence of GuardVPO’s incremental contribution across multiple update stages.

Table B. Held-out performance with a 128-episode GuardVPO update interval.

SettingUOR ↓TSR ↑SUCR ↑
H0+G0H_0+G_023.00%34.00%30.33%
H128+G1H_{128}+G_116.33%42.00%37.67%
H256+G2H_{256}+G_215.67%45.00%39.33%
H384+G3H_{384}+G_314.33%35.67%30.33%
H512+G3H_{512}+G_312.33%39.00%34.33%
H512+G4H_{512}+G_410.67%45.00%40.33%

We acknowledge the modest gains on the original external Guard benchmarks. Limited evaluation coverage and the strong initial capabilities of AgentDoG1.5-Unified-Qwen3.5-4B and SingGuard-NSFA-4B may contribute to this observation. We therefore evaluate an alternative Guard, ShieldAgent, fine-tuned from Qwen2.5-7B-Instruct, on the held-out test set and two out-of-distribution (OOD) benchmarks, R-Judge and ATBench.

As shown in Table C, updating ShieldAgent at fixed H512H_{512} reduces UOR by 5.00 percentage points and increases TSR/SUCR by 3.67/5.00 points, supporting GuardVPO's incremental benefits with another backbone.

Table C. ShieldAgent held-out performance before and after GuardVPO with S-Harness fixed.

SettingUOR ↓TSR ↑SUCR ↑
H512+G0H_{512}+G_021.67%35.00%27.00%
H512+G1H_{512}+G_116.67%38.67%32.00%

Table D shows Unsafe F1/Recall gains of 0.17/3.39 percentage points on R-Judge and 6.59/10.06 points on ATBench. These results support improved unsafe-event detection on external benchmarks, while lower Safe Specificity reveals a trade-off between unsafe-event sensitivity and safe-class specificity.

Table D. ShieldAgent on external benchmarks before and after GuardVPO. F1 and Recall refer to the unsafe class.

GuardR-Judge F1 ↑Recall ↑SafeSpec ↑ATBench F1 ↑Recall ↑SafeSpec ↑
ShieldAgent83.39%89.56%71.91%51.90%41.25%82.50%
ShieldAgent + GuardVPO83.56%92.95%67.16%58.49%51.31%76.14%

These experiments support GuardVPO's complementary contribution across evolution stages and Guard backbones. We will incorporate these additional results and analyses into the revised manuscript.

AuthorsOpenReview ↗

Response to Question 1: GuardVPO Update Interval and Longer-Stream Evaluation

Thank you for this valuable suggestion. We acknowledge that the original 512-episode GuardVPO update interval allows relatively few Guard updates, limiting our observations of the adaptation process across different update frequencies. We would like to clarify that 512 episodes is a configurable experimental parameter, rather than a fixed constraint of the SafeCoEvo framework.

In the original experiments, we chose this interval to accumulate sufficient event-level safety experience before each parameter update while balancing update frequency, optimization stability, and computational cost. This interval is not an inherent requirement of the framework and can be adjusted according to the deployment setting and the accumulation of runtime experience. Throughout the continual evolution stream, S-Harness updates after each episode, while GuardVPO periodically updates the Guard parameters using all event-level safety experience accumulated up to that point.

Using Qwen3.7-Flash as the task-performing model, we conduct additional experiments with update intervals of 128 and 256 episodes and compare them with the original 512-episode setting.

Table A. Held-out performance under different GuardVPO update intervals.

GuardVPO update intervalUOR ↓TSR ↑SUCR ↑
128 episodes10.67%45.00%40.33%
256 episodes12.33%55.00%50.33%
512 episodes10.67%55.00%49.33%

As shown in Table A, the 512-episode interval achieves the joint-lowest UOR (10.67%) and a TSR of 55.00%, matching the 256-episode setting, while its SUCR is slightly lower (49.33% vs. 50.33%). Under the evaluated setting, these results support 512 episodes as a reasonable default that achieves competitive safety and task utility with fewer Guard updates. The comparison also indicates that increasing the update frequency does not necessarily improve final performance.

We further extend the continual evolution stream to 2,048 episodes while retaining the 512-episode GuardVPO update interval, allowing us to evaluate the framework over a longer operating horizon. The results are reported in Table B.

Table B. Performance comparison over a continual evolution stream of 2,048 episodes.

MethodUOR ↓TSR ↑SUCR ↑
SHE47.85%30.76%24.71%
SafeCoEvo (Static)28.52%44.34%35.64%
SafeCoEvo20.56%53.52%44.24%

As shown in Table B, compared with the static configuration, SafeCoEvo reduces UOR by 7.96 percentage points and increases TSR and SUCR by 9.18 and 8.59 percentage points, respectively. These results show that SafeCoEvo's co-evolution mechanism continues to improve both safety and task utility over a longer task stream, complementing the evaluation across different update intervals.

Together, these additional experiments provide a more comprehensive assessment of GuardVPO under different update frequencies and of SafeCoEvo over a longer evolution horizon. We have added the update-interval comparisons to Appendix G and the longer-stream evaluation to Appendix H. Appendix G also explains the configurability of the update interval and the rationale for the default 512-episode setting, while the revised Continual Evaluation Protocol in Section 4.1 clarifies the actual update procedure. These additions address the concern about the limited number of Guard updates and provide a clearer account of SafeCoEvo’s continual adaptation process.

AuthorsOpenReview ↗

Response to Question 2: Joint Adaptation of S-Harness and GuardVPO During Continual Evolution

Thank you for pointing this out. We sincerely apologize for the confusion caused by the imprecise description of the experimental protocol in Section 4.1. We have revised the Continual Evaluation Protocol paragraph in Section 4.1 to clarify the actual procedure.

In the original experiments, S-Harness is updated after each episode throughout the entire 1,024-episode continual evolution stream. GuardVPO updates the Guard parameters every 512 episodes using event-level safety experience accumulated from completed episodes. Specifically, the initial Guard is used during the first 512 episodes while S-Harness continually adapts. After episode 512, GuardVPO performs its first parameter update, and the updated Guard is deployed during episodes 513–1,024, while S-Harness continues to update after each episode. After episode 1,024, GuardVPO updates the Guard again, and the resulting configuration is evaluated on an independent held-out test set. Thus, the original experiments already incorporate joint continual adaptation of both components within the same task stream, with each component operating at a different update frequency.

Thus, S-Harness and GuardVPO co-evolve on the same continual evolution stream at different timescales. S-Harness undergoes rapid, episode-by-episode updates, while GuardVPO performs periodic parameter updates. We will clarify the wording that may have suggested that the two components evolve separately. We again apologize for the confusion caused by the original description and will ensure that Section 4.1 accurately explains the experimental procedure in the revised manuscript.

We thank the reviewer again for the careful reading and constructive feedback. We hope that this clarification and the corresponding revisions provide a clearer understanding of SafeCoEvo’s experimental design and core contribution, as well as a stronger basis for reassessing our work.

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

This paper studies continual test-time safety adaptation for agents. We ask how a system can use past interactions and feedback already received to improve safety on subsequent unseen tasks as tasks arrive sequentially. To the best of our knowledge, SafeCoEvo is the first agent safety framework to co-evolve a Harness and a Guard at test time. After each task, S-Harness converts recent experience into explicit, reusable safety knowledge. GuardVPO periodically updates the Guard using accumulated event-level experience so that it can apply what it has learned about risk to new situations. On the continual evolution stream, SafeCoEvo reduces the unsafe outcome rate (UOR) by 10.05 percentage points compared with SHE, the strongest evolution baseline, and increases the task success rate (TSR) and the safe and useful completion rate (SUCR) by 12.15 and 15.67 percentage points, respectively. These results show that updating S-Harness and the Guard at different frequencies on the same task stream can improve both safety and task utility while the task-performing model remains frozen.

During the rebuttal period, we conducted additional experiments in response to constructive suggestions from the three reviewers.

  • GuardVPO update intervals and contributions at different stages (AcnG). We evaluated Guard updates every 128 and 256 episodes and compared these intervals with the original 512-episode interval. Comparisons before and after individual Guard updates, with S-Harness held fixed, showed that updating the Guard provided additional benefits at multiple stages of evolution (Appendix G).

  • Ablation of the five S-Harness components (Vu6M). With GuardVPO disabled, we froze updates to each of the five components in turn. Freezing any one component increased UOR and decreased both TSR and SUCR, indicating that the five components play complementary roles in continual adaptation (Section 4.6, Table 7).

  • Evaluation with an alternative Guard backbone (AcnG, Vu6M, 5Fqx). We replaced the original Guard with ShieldAgent and evaluated the contributions of S-Harness evolution and GuardVPO updates separately. We also compared ShieldAgent’s risk assessments before and after the GuardVPO update on R-Judge and ATBench. Recall for unsafe events improved on both benchmarks (Appendix D.1).

  • Evaluation across task-performing models (5Fqx). We added experiments using GPT-6.1-terra as the task-performing model and compared its performance with that of DeepSeek-v4.1-Flash and Qwen3.7-Flash under the same safety framework. The results show that SafeCoEvo’s benefits extend across task-performing models (Appendix E).

  • Cross-benchmark generalization (Vu6M, 5Fqx). To address the risk of overfitting when adaptation and testing use the same source benchmarks, we froze the evolved S-Harness and Guard and evaluated them on selected scenarios from AgentDojo and AgentDyn. SafeCoEvo achieved a higher safe completion rate than SHE on both external benchmarks (Appendix C.1).

  • Robustness across random seeds (5Fqx). We repeated the experiments using three independent random seeds and reported the mean, sample standard deviation, and 95% confidence interval for the mean. Across all three seeds, SafeCoEvo achieved lower UOR and higher TSR and SUCR than SHE (Appendix F).

  • Longer task streams and rule conflicts (AcnG, 5Fqx). We extended the task stream to 2,048 episodes while retaining the 512-episode Guard update interval. Over the full stream, SafeCoEvo continued to outperform the static configuration in both safety and task utility (Appendix H). In response to 5Fqx’s follow-up question about rule conflicts, we also presented two examples from the evolution logs showing how S-Harness refines the conditions under which existing rules apply and incorporates revised guidance into reusable skills.

We thank the Area Chair and all three reviewers for the time they devoted to reviewing the paper and participating in the discussion. We believe we have responded fully to the concerns raised by all three reviewers and supported our responses with evidence. Your suggestions helped us explain more clearly how S-Harness and GuardVPO complement each other during co-evolution. The component-wise ablations, cross-model and cross-benchmark evaluations, and longer-stream experiments also provide stronger empirical support for the paper’s central findings. We believe these additions have substantially improved the clarity of the paper and the completeness of its experimental evaluation. We respectfully ask the Area Chair and reviewers to give these improvements and the paper’s contribution to agent safety research full consideration in the final assessment.

The authors of submission 31378

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.