acceptodds
Withdrawn from ICLR 2027

PUARobot: Psychological Manipulation for Jailbreaking Embodied LLM Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Embodied AI agents increasingly rely on large language models (LLMs) for language understanding and task planning in real-world tasks, yet LLM jailbreak vulnerabilities may translate harmful text into dangerous physical actions. Inspired by psychological manipulation in human interactions, we ask whether similar social, emotional, and cognitive pressure can bypass embodied-agent safeguards and induce unsafe actions. In this work, we investigate psychological manipulation-based jailbreak attacks on embodied agents. Specifically, we (1) propose PUARobot, a black-box jailbreak attack framework that systematizes seven psychological-manipulation strategies as attack methods to reveal the safety weaknesses of embodied agents under different forms of manipulative pressure; (2) establish a paradigm that transforms a direct single-turn request into an attacker-authored, multi-turn, psychological-manipulation-based disguised dialogue, gradually applying pressure in the context to induce unsafe action plans; and (3) develop an action-centered embodied-safety evaluation mechanism and feed evaluation feedback into dialogue search to iteratively optimize attack prompts, addressing the tendency of existing benchmarks to overemphasize text and misjudge unsafe actions. Experiments across multiple open- and closed-source models show that PUARobot achieves an average Attack Success Rate (ASR) of 76.2%, compared with 7.3% for the strongest BadRobot baseline. Further experiments on Code as Policies, ProgPrompt, and VisProg demonstrate its effectiveness across mainstream embodied AI frameworks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Robotics, Safety Risks, Embodied AI, LLM, Agent, Jailbreak attack

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_8oQ7OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PUARobot, a framework for inducing embodied LLM planners to generate harmful actions through natural-language attacks. It turns seven forms of social and psychological pressure into fabricated dialogues, submits each complete dialogue as a single user input, and iteratively searches using action scores and API-validity feedback. Here, “multi-turn” refers to the input structure; the target model does not participate in the preceding exchanges (§3.1, §3.3). Across 321 tasks and 8 target models, with up to 50 queries per task, the authors report an average attack success rate of 76.2%, compared with 7.3% for the strongest BadRobot baseline (§4.1, Table 1). Table 3 provides component ablations, while §4.5 and Appendix F present simulation and visual-program execution examples.

Critical strengths
  1. The evaluation addresses a meaningful distinction in embodied safety. I appreciate that the method separates action-harm scoring from format checks against declared APIs and evaluates only newly generated target outputs, without counting attacker-authored historical actions as target behavior (§3.3–3.4, Appendix C). This helps distinguish a dangerous description from a dangerous plan that can be passed to an executor. The paper also explicitly notes that interface validity does not establish successful physical execution.
  2. The experiments offer broad coverage and component analysis. Table 1 covers 8 models and 7 risk categories; Table 3 compares single-turn inputs, fabricated dialogues, and the addition of action feedback. For example, Commitment Escalation improves from 0.35 to 0.56 and then to 0.65. I also value the evaluator checks with final human-confirmed judgments in Table 2 and the execution examples in §4.5 and Appendix F. Together, these results support the attack's effectiveness in the tested settings, although the examples themselves do not estimate execution success rates.
Critical weaknesses
  1. The main success metric excludes the central failure mode of verbal refusal accompanied by harmful actions. Eq. (2) requires an action score of at least 4, valid APIs, and a language score of at least 3. Appendix C assigns explicit refusals and warnings scores of 0–2. A substantively harmful plan that passes the interface checks is therefore counted as an attack failure if accompanied by refusal text. I find this inconsistent with the action risk emphasized in §3.4. It may particularly underestimate BadRobot's Safety Misalignment baseline, which specifically targets discrepancies between language and actions (Appendix D.1). Until success is recomputed using harmful actions and API validity, with language behavior reported separately, I do not think the 76.2% versus 7.3% gap adequately establishes the difference in action-level safety.
  2. The controls do not establish the separate contributions of psychological pressure and adaptive search. The main comparisons use fixed BadRobot templates. I recognize that Table 3 places four conventional strategies in dialogue and feedback settings, but it does not compare against existing automated persuasive attacks or adaptive attacks on robots. Relevant examples include PAP and RoboPAIR, both research directions already cited in the paper. On GPT-5.5, the full system's 0.71 versus 0.65 for the best fixed strategy also does not by itself establish that Thompson sampling outperforms random strategy selection (Table 1, Table 3, Appendix D.3). Without these controls, I cannot tell how much of the gain comes from the specific design choices emphasized here, rather than strategy diversity, richer context, or general feedback-based search.
Questions for authors
  1. Clarification and reanalysis: Why does action-level attack success also require a language score of at least 3? Using the existing outputs, please report success based on harmful, API-valid actions separately from success that additionally requires verbal compliance, along with each method's rate of harmful actions accompanied by verbal refusal. Does the gap between BadRobot and PUARobot remain?
  2. Additional controls: How does PUARobot compare with RoboPAIR or PAIR-style adaptive attacks under the same target models, interfaces, query limits, and evaluation rules? How much advantage remains when PAP-style prompts or conventional strategies use the same action-feedback search process? My main interest here is in establishing the incremental contribution, rather than expanding the task set.
  3. Additional evidence: How do success rates and query costs change when Thompson sampling is replaced with uniform random strategy selection? Since category statistics are shared across tasks in the same category, please specify task ordering and reset rules for each independent run, and report variation across repeated runs. This directly affects the interpretation of the improvement from 0.65 to 0.71.
Reviewer_y8GpOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PUARobot, a black-box jailbreak framework for embodied LLM agents based on psychological manipulation. The method introduces seven psychological manipulation strategies, constructs attacker-authored multi-turn dialogues to induce harmful action plans, and uses an action-centered evaluator with feedback-driven adaptive search to improve attack effectiveness. The authors evaluate PUARobot on 321 harmful instructions across eight LLMs and seven risk categories, reporting an average attack success rate (ASR) of 76.2%, compared with 7.3% for the strongest BadRobot baseline. Additional experiments demonstrate selected attack outcomes using Code as Policies, ProgPrompt, and VisProg.

Critical strengths
  1. Important research problem. The paper investigates jailbreak vulnerabilities in embodied LLM agents, where unsafe model outputs may lead to harmful physical actions. This is an important problem for the safety and reliability of robotic systems.

  2. Comprehensive attack framework. The proposed method combines psychological manipulation strategies, synthetic dialogue construction, action-centered evaluation, and feedback-driven optimization into a unified attack pipeline. The explicit consideration of action-interface validity is useful for evaluating attacks against LLM-based robotic planners.

  3. Extensive LLM evaluation and useful ablation studies. The authors evaluate the method across eight LLMs and seven categories of harmful tasks, demonstrating substantial improvements over the selected baselines. Table 3 further examines psychological manipulation strategies, disguised dialogues, and feedback-driven optimization, providing evidence for the effectiveness of these components.

Critical weaknesses
  1. Limited embodied-specific novelty. Most components of PUARobot, including psychological manipulation, synthetic dialogues, and adaptive search, are general LLM jailbreak techniques. Beyond API validity checking, the method does not explicitly incorporate physical constraints, visual observations, or closed-loop interactions. It remains unclear what technical challenges specific to embodied AI are addressed.

  2. Limited evaluation on modern embodied AI systems. The experiments rely on earlier LLM-based frameworks (Code as Policies, ProgPrompt, and VisProg), without evaluating recent vision-language-action (VLA) models. Since VLA models differ substantially in their input and action representations, the generalizability of PUARobot to modern embodied AI systems remains unclear.

  3. Insufficient explanation of the effectiveness of psychological manipulation. Although Table 3 shows that psychological manipulation strategies outperform classic jailbreak strategies, the underlying reasons remain unclear. The improvements could also result from richer context, fabricated assistant responses, or prompt structure rather than psychological manipulation itself. More controlled comparisons are needed to isolate these effects.

Questions for authors

1.What aspects of PUARobot specifically address embodied AI challenges? How would the method differ when applied to general-purpose LLMs or other tool-using agents?

2.Have the authors considered evaluating PUARobot on recent VLA models? How would the proposed approach transfer to models without conventional conversational interfaces or structured action APIs?

3.Can the authors provide controlled experiments that isolate the effects of psychological manipulation from prompt structure, contextual information, and fabricated assistant responses, as the naive single turn PUA attacks outperform Badrobot and we are curious about the concrete reasons here.

4.How would the BadRobot baselines perform under the same adaptive dialogue construction and feedback-driven optimization pipeline used by PUARobot?

Reviewer_2YsPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PUARobot, a black-box jailbreak attack framework targeting large language model (LLM)-driven embodied agents. The method incorporates seven psychological manipulation strategies to construct attacker-authored, multi-turn disguised dialogues that gradually induce target models to generate unsafe action plans. Furthermore, the authors integrate Thompson Sampling-based strategy selection, historical feedback retrieval, and an action-centered hybrid evaluation mechanism to enable adaptive attack optimization. Experiments are conducted on 321 harmful instructions across seven risk categories and eight target LLMs, with additional simulation demonstrations using Code as Policies, ProgPrompt, and VisProg. The results show that PUARobot achieves an average Attack Success Rate (ASR) of 76.2%, substantially outperforming the existing BadRobot baselines.

Critical strengths
  1. A well-structured adaptive attack framework. The paper integrates psychological manipulation strategies, multi-turn disguised dialogue construction, action-level feedback, and historical experience retrieval into a unified iterative optimization framework. Compared with fixed attack templates, this design allows the attack process to adapt to responses from different target models, improving the flexibility and effectiveness of the proposed approach.

  2. Extensive experiments across multiple models and embodied AI frameworks. The authors evaluate PUARobot on eight target LLMs and seven categories of safety risks. Additional experiments using Code as Policies, ProgPrompt, and VisProg demonstrate the effects of the generated action plans across different downstream frameworks. Furthermore, the ablation studies investigate the contributions of individual psychological manipulation strategies, disguised dialogues, and feedback mechanisms, providing empirical support for the effectiveness of the proposed approach.

Critical weaknesses
  1. Inconsistency between the attack success criterion and the action-centered safety evaluation principle. The paper emphasizes that safety risks in embodied agents should primarily be determined by their generated actions rather than their natural language responses. However, Equation (2) additionally requires the language score to satisfy stl≥3s_t^l \geq 3, alongside the action harmfulness score and interface validity. Consequently, even if a model explicitly refuses a harmful request in its language response while generating a valid and harmful action sequence, the attempt is still classified as unsuccessful. This appears inconsistent with the paper's emphasis on language–action safety misalignment and may underestimate action-level safety violations. The authors are encouraged to report Action-Only ASR, Language Compliance Rate, and joint attack success separately, and to clarify why language compliance is a necessary condition in an action-centered evaluation framework.

  2. Numerical inconsistencies in evaluator validation raise concerns about the reliability of the experimental results. Section 4.3 validates Qwen3-Plus as the attack evaluator and reports Accuracy, Precision, and F1 in Table 2. However, some reported metric combinations appear inconsistent with the standard binary classification definition of F1. For example, GPT-5.5 achieves a Precision of 91.85% and an F1 score of 95.75%. Under the standard F1 formula, this combination would require a Recall exceeding 100%, which is mathematically impossible if both metrics are computed from the same confusion matrix. Since the evaluator is used both to guide attack optimization and to determine final attack success, such inconsistencies raise concerns about the reliability of the reported results. The authors should carefully verify these metrics, provide Recall and confusion matrices, clarify the calculation procedures, and further examine the agreement between evaluator predictions and human annotations.

  3. Inconsistencies between simulation experiments and their descriptions in the appendix. Some descriptions of the simulation experiments do not match the corresponding tasks. For example, Appendix F.2 discusses a stove-related task but describes the agent approaching a bookshelf, locating a folder, and manipulating private household objects. Such inconsistencies make it difficult to establish a clear correspondence between the task definitions, execution trajectories, and reported results. The authors should carefully review the simulation experiments and ensure that the task descriptions, environment configurations, execution steps, and visual evidence are consistent throughout the paper.

Questions for authors

Q1. Could the authors clarify why language compliance is required for attack success in an action-centered evaluation framework? According to Equation (2), an attack is considered successful only when the language score satisfies stl≥3s_t^l \geq 3, even if the generated action sequence is harmful and API-valid. This appears to exclude cases where an agent verbally refuses a harmful request but still produces unsafe actions. Could the authors explain the rationale behind this requirement and provide additional results using an action-only success criterion?

Q2. Could the authors clarify the calculation of the evaluator validation metrics reported in Table 2? Some combinations of Precision and F1 appear inconsistent with the standard binary classification definitions. For instance, GPT-5.5 achieves a Precision of 91.85% and an F1 score of 95.75%, which would require a Recall greater than 100% if both metrics were computed from the same confusion matrix. Could the authors verify these results and provide the corresponding Recall values or confusion matrices to clarify the evaluation procedure?

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

Related papers

Discussion (0)

Sign in to comment.