acceptodds
Under review as a conference paper at ICLR 2027

SEAD: A State-Based Perspective on Attack and Defense in Tool-Using Agents

Abstract

Language-model agents increasingly use tools to act on external systems. Earlier actions can alter files, permissions, database records, or other state, making a later routine-looking action harmful. Yet the visible interaction may not reveal the underlying state needed to assess that action. We formulate attack and defense as partially observed state control in SEAD, deriving their design requirements from this shared execution process. Because attackers supply instructions while the target chooses concrete actions, DART decomposes harmful goals into locally plausible steps and uses feedback from actual tool execution to guide trajectory search. The defender must decide before execution with incomplete state evidence. SAGE can therefore investigate relevant state through read-only queries before allowing or blocking each action, including those proposed after a block. We construct an environment-verifiable dataset integrating controlled initial states, replayable tool environments, and task-specific executable checks. Across four target models, DART improves semantic attack success by 18.8–35.9 percentage points over the competing baseline, with consistent gains under executable verification. On recorded trajectories, SAGE preserves 95.79% of benign trajectories while intercepting 92.73% of harmful paths by the harm-enabling boundary. In online attack-defense evaluation, it reduces DART’s executable attack success from 48.0% to 4.0%. SAGE remains effective across four attack methods and generalizes to out-of-domain environments. Our code and data is available at this anonymous repository: https://anonymous.4open.science/r/SEAD-review-2027.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Agentic Defense, Red-Teaming, AI Safety

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_WJGKOpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper introduces SEAD, a unified framework that models attack and defense in tool-using agents as a partially observed state-control problem, emphasizing how cumulative environment changes can lead to harmful outcomes. Building on this formulation, the authors propose DART, an adaptive attack method that uses execution feedback and tree search to construct multi-step attacks, and SAGE, a pre-execution defense that actively investigates environment states before allowing or blocking tool actions.

Experiments cover 187 harmful tasks across four target models, with additional offline and online defense evaluations. DART consistently outperforms existing attack baselines, while SAGE reduces executable attack success from 48.0% to 4.0% in online evaluation and preserves 95.79% of benign trajectories in offline evaluation. The authors also report encouraging generalization to held-out tool domains and different attack methods.

Critical strengths
  1. A useful state-based perspective on agent safety. The paper makes a convincing case that harmfulness should be assessed through actual environment transitions rather than isolated tool calls. The unified formulation of attack and defense is intuitive and practically relevant.

  2. Strong empirical results with executable verification. The use of replayable environments and executable success checks is a meaningful strength. In particular, SAGE reduces online hard ASR from 48.0% to 4.0%, with encouraging results on held-out domains.

Critical weaknesses
  1. The attack comparisons are not sufficiently controlled. DART uses task-specific executable verifier progress during search, and its feedback critic receives completion criteria (Sec. 4.1, Appendix F.3). The competing methods also operate under different search budgets. Without matched feedback access and computational budgets, it is difficult to attribute DART's gains to the proposed search strategy.

  2. The contribution of state investigation is not adequately isolated. SAGE lacks a controlled ablation without environment queries. Moreover, 58.18% of harmful trajectories are blocked before the annotated closure, and online benign task completion is not evaluated. These limitations weaken the claims about precise interception and safety–utility balance.

Questions for authors
  1. How much of DART's improvement remains under matched execution budgets and without privileged verifier-progress feedback?

  2. How does SAGE perform without read-only environment queries, keeping the backbone and training setup unchanged? Could the authors also clarify how the harmful-closure annotations are validated and report online benign task completion?

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

This paper presents SEAD, a unified framework for studying attacks and defenses in tool-using agents from a state-based perspective. The key observation is that individually benign-looking actions can accumulate changes to the environment and eventually lead to harmful outcomes. Based on this formulation, the authors introduce DART, an adaptive attack method that decomposes harmful objectives into intermediate steps and uses execution feedback and tree search to guide subsequent attacks. They also propose SAGE, a pre-execution defense that can actively inspect environment state through read-only queries before allowing or blocking a tool action. The authors evaluate DART on different tasks across four target models, showing improvements over existing attack methods under both semantic and executable success criteria.

Critical strengths

1.Strong and well-motivated problem. The paper addresses an important issue in tool-using agent safety. The motivation is convincing, as harmful outcomes may depend on accumulated environment states rather than individual actions.

2.Strong empirical performance. The proposed methods achieve substantial improvements over existing baselines. The results are supported by various settings.

Critical weaknesses

1.The main motivation is not fully supported by the experiments. The paper emphasizes how accumulated state changes can make later actions harmful. However, most experiments focus on final attack success, without directly examining whether the preceding state changes are actually necessary for the harmful outcome.

2.The utility evaluation is somewhat limited. The paper measures utility based on whether the defender allows all actions in recorded benign trajectories, rather than whether agents can complete legitimate tasks during online execution. Also, SAGE blocks actions earlier than the harmful boundary in most attack trajectories, raising questions about how well it preserves useful intermediate actions.

Questions for authors

1.The paper emphasizes that accumulated state changes can make later actions harmful. Can the authors provide evidence that these earlier state changes are actually necessary for the harmful outcome, rather than the attack simply succeeding through multi-turn instructions?

2.Table 3 shows that SAGE blocks actions before the harmful boundary in 58.18% of attack trajectories. How often do these early blocks interrupt otherwise useful actions? Also, could the authors evaluate benign task completion in an online setting rather than only measuring whether recorded actions are allowed?

3.The attack methods use different search budgets and feedback mechanisms, which makes their performance somewhat difficult to compare. Could the authors clarify how these differences affect the results? In particular, how much does DART benefit from the intermediate executable progress feedback?

2 comments

Thanks for the detail rebuttal. I will maintain the current positive score.

We are grateful to the reviewer for the positive assessment of our contributions and empirical results, and for taking the time to consider our rebuttal. We would be happy to answer any follow-up questions.

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

This paper studies attacks and defenses for tool-using agents from the perspective of changes to the environment state. It proposes DART, which searches over executed attack trajectories using model assessments and executable progress feedback, and SAGE, which can issue private, read-only queries before allowing or blocking a tool action. The evaluation covers 187 malicious tasks across four domains. DART improves attack success across four target models, while SAGE is evaluated on recorded trajectories and in online interactions where the attacker can respond to blocked actions.

Critical strengths
  1. The paper focuses on a real difficulty in agent safety: the effect of a tool call can depend on state that is missing from the visible conversation. The examples make this motivation easy to follow.
  2. The evaluation includes executable outcome checks and online interactions after a defense intervenes. Continuing the interaction after a block is particularly useful, since an initial refusal need not prevent the attacker from finding another route.
  3. The reported improvements seems to be substantial. DART improves Hard ASR across all four targets, and SAGE reduces DART’s online Hard ASR from 48.0% to 4.0% on the 75-task subset. These results deserve attention, although I have concerns about what the comparisons establish.
Critical weaknesses
  1. DART uses task-specific verifier feedback during search, while the baselines retain their original feedback interfaces. Without matched feedback and execution budgets, it is difficult to attribute the gains to DART’s search design.
  2. SAGE combines additional training with access to investigation tools. A control trained on the same data without queries would help isolate the benefit of investigation. A comparison with SafeMCP also seems relevant.
  3. The timing metrics favor allowing actions before the designated goal-completion point. Some earlier actions may already be harmful, and preserving more of an attack trajectory does not necessarily preserve useful benign work.
  4. Semantic and Hard ASR differ substantially, yet the offline defense evaluation uses Semantic-success trajectories. The paper should explain these disagreements and report how many evaluated harmful paths also pass executable verification.
Questions for authors
  1. Do the attack baselines receive the same executable progress feedback as DART? What are the target-call and candidate-execution budgets for each method, and how does DART perform without verifier feedback?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.