acceptodds
Under review as a conference paper at ICLR 2027

Probing Biosafety Risks in LLMs via Skill-Driven Agentic Jailbreaking

Abstract

Large language models (LLMs) are increasingly used in biological research, raising growing concerns about their potential biosafety risks. However, existing safety evaluations largely rely on static, single-turn queries, which may substantially underestimate risks that emerge through sustained and adaptive interaction. We introduce BioSafeProbe, a skill-driven agentic jailbreaking framework for systematically probing such biosafety risks in LLMs. BioSafeProbe formulates safety testing as a sequential decision process that adaptively explores model safety boundaries over complete interaction trajectories. It decomposes complex risk objectives, dynamically selects and orchestrates reusable interaction skills according to the evolving testing state, evaluates intermediate risk exposure and task progress, and updates subsequent strategies based on model feedback and prior interaction experience. We further construct BioRisk, an open-ended biosafety evaluation dataset comprising 300 risk tasks across 10 categories. Experiments across 10 frontier LLMs show that BioSafeProbe increases the average risk exposure rate from 31.3% to 87.7%, revealing substantially more biosafety risks that would otherwise remain undetected. Fine-grained analyses further show that some risks emerge only after multiple interaction turns and exhibit distinct patterns across risk categories, target models, and interaction trajectories.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, AI Safety, Biosafety, Red Teaming

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_r7fuOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces BioSafeProbe, an agentic framework for evaluating biosafety risks through adaptive multi-turn interactions, together with BioRisk, a dataset of 300 tasks across ten risk categories. The framework selects interaction skills using model feedback and historical experience. Experiments report increased risk exposure compared with direct queries and several jailbreak baselines.

Critical strengths
  1. Evaluating safety under sustained interaction addresses a meaningful limitation of single-turn testing and is relevant to understanding model behavior in realistic usage.

  2. The study evaluates ten target models on a shared task set and includes domain-expert task screening, component ablations, and analyses across categories and interaction turns.

Critical weaknesses
  1. Dataset coverage and risk measurements are insufficiently validated. BioRisk is expanded from one seed per category, but the paper provides limited evidence about category coverage, within-category diversity, or annotation agreement. The reported expert screening is useful, but it concerns task validity rather than the accuracy of model-response risk scores. StrongREJECT alone does not establish that the reported scores reliably capture biological correctness or domain-specific risk severity.

  2. The most relevant multi-turn comparison is missing. The paper acknowledges prior scientific safety evaluation using Crescendo in SoSBench, yet does not include a corresponding comparison. The selected baselines provide useful reference points, but do not sufficiently establish the contribution beyond existing multi-turn evaluation approaches. Moreover, baseline comparisons cover only two of the ten target models.

  3. The experiments do not adequately isolate the source of the gains. The main comparison changes both the evaluation procedure and the interaction budget, from one direct query to up to twenty adaptive turns. Matching method-specific iteration counts does not necessarily match model calls or computational cost. In addition, the trajectory metrics aggregate the highest-scoring individual response, providing limited evidence about risks arising specifically from information accumulated across turns.

Questions for authors
  1. What evidence supports the coverage and diversity of BioRisk, and how consistently did the experts agree? Have the model-response risk scores been independently checked by domain experts?

  2. How does BioSafeProbe compare with closely related multi-turn evaluation approaches, particularly the Crescendo setting used in SoSBench, under comparable evaluation budgets?

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

This paper investigates biosafety risks in large language models (LLMs). Unlike conventional evaluations that rely on a single prompt to assess whether a model refuses potentially hazardous requests, this paper employs an LLM agent to conduct adaptive, multi-turn testing. The agent adjusts its testing strategy in response to the model’s replies, enabling it to uncover risks that may remain undetected in single-turn evaluations. To this end, it introduces BioSafeProbe, a benchmarking framework that dynamically selects probing skills and adapts its strategy based on model feedback. It then uses the BioRisk benchmark to evaluate 10 LLMs under both single-turn and multi-turn settings. The results show that the average risk exposure rate increases from 31.3% in single-turn tests to 87.7% in multi-turn interactions, highlighting the limitations of single-turn evaluations in assessing LLM biosafety.

Critical strengths

S1. This paper investigates biosafety risks in LLMs through adaptive multi-turn interactions. I find this approach interesting, as single-turn evaluations may fail to capture risks that emerge over the course of multi-turn conversations. The paper employs an LLM-based agent that dynamically adjusts its testing strategy based on the target model's responses, which seems to be a reasonable approach for identifying such risks.

S2. The experiments cover 10 LLMs, 300 tasks, and 10 biosafety risk categories, providing a relatively comprehensive evaluation. The results further demonstrate that multi-turn testing can uncover risks that may be overlooked by single-turn evaluations.

Critical weaknesses

W1. The risk evaluation relies heavily on an automated evaluator. The paper uses StrongREJECT scores and a fixed threshold to determine whether a task exposes a biosafety risk. However, there is limited human evaluation to validate these results. A high risk exposure rate does not necessarily indicate that the model generates accurate or practically useful harmful information. This makes it difficult to assess the actual severity of the reported risks.

W2. The comparison with existing jailbreak methods is relatively limited. While the main experiments cover 10 LLMs, the comparisons with PAIR, GPTFuzzer, and ReNeLLM are conducted on only two target models. In addition, the paper does not provide a detailed analysis of token usage or API costs. As a result, it remains unclear whether the proposed method can achieve similar improvements across other models or whether these improvements justify the additional computational overhead.

Questions for authors
  1. How reliable is StrongREJECT in evaluating biosafety risks? Can human evaluation confirm that the identified responses contain scientifically valid and harmful information?
  2. Why are the comparisons with existing jailbreak methods conducted on only two LLMs? Can the results generalize to the other evaluated models?
  3. How does BioSafeProbe compare with existing methods in terms of token usage, API costs, and overall efficiency?
Reviewer_pP3yOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This work focus on bio-risk probing of current frontier LLM. It introduces BioRisk, a new evaluation dataset with 300 cases to assesse biosafety risks with fine-grained categories. And it proposes BioSafeProbe, which is a skill-driven agentic framework for multi-turn jailbreak probing. With this new framework, the evaluation run across 10 LLMs, compared to existing baselines, shows the effectiveness of BioSafeProbe and unreveal the hidden vulnerabilities beyond static evaluations.

Critical strengths
  • This work construct a new evaluation dataset BioRisk.
  • This work propose a new agentic skill-based multi-turn jailbreak BioSafeProbe.
  • The broad evaluation on 10 LLMs and compared baselines reveal the hidden biorisk vulnealiblity beyond static evaluation.
Critical weaknesses
  • It is unclear why the new constructed dataset BioRisk resolve the gap from existing evaluation benchmark. For example, I did not clear distinguishment the novelty of produced BioRisk compared to SoSBench, which covers also bio risk and contains also open-ended scientific misuse prompts. Therefore, the existing benchmark is natually the testbed for BioRisk already.
  • Evaluation is not robust.
    • The result metric all depends on StrongReject judge. First, the details of this judge is not provided at all. It is unclear whether original GPT4 version or fine-tuned gemma model is used. Secondly, it is unclearly how this judge performs towards biorisk task and multiturn setup, especially the alignment with human judge.
    • Also the refinement uses the same judge during attack and evaluation stage.
  • The design lacks enough novelty. The overall design seems just a ensemble exsiting prompt-based jailbreak strategy, it is unclear why the approach is novel compared to existing approach, except for a 'skill-based' framing. Also, the connection of this method to the bio-risk is missing. In my understanding, it simply applied the approach just on this domain, but actually not specific to Bio domain.
  • Experiment is not solid.
  • The baseline is very limited. Some more recent and relevant attack approach should be included.
  • As mentioned above, the approach is not for BioRisk dataset only, it should test on other benchmark.
Questions for authors
  • In the design of BioRisk dataset, why does it need to contain Goal–Behavior–Target, and what is the distinguishment among these?
  • I don't understand the exact meaning of RER metric.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.