acceptodds
Under review as a conference paper at ICLR 2027

Does Hand-Designed Trajectory State Improve Pre-Execution Control?

Abstract

We test whether accumulating risk across an agent’s trajectory improves prevention over evaluating each action independently. Across separately analyzed live model studies and released AgentDojo traces, the evaluated hand-designed state establishes no prevention advantage over a stateless per-action rule. Replaying frozen calibration grids raises trajectory prevention to the stateless level but not above it. The live attack studies share seven authored scenarios, and we do not evaluate learned sequence policies. A diagnostic identifies why one apparent observability failure occurred: the controller discarded a tool-name signal already available before execution. Separately, removing returned tool content causes model evaluators to misidentify injection origin in 33 of 40 blinded packets. Within these evaluations, accumulated risk does not improve prevention, whereas retained content supports incident reconstruction. The resulting design lesson is to preserve decision-relevant information before compressing it into a risk state, and to evaluate prevention separately from reconstruction.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI agent safety, prompt injection, pre-execution control, trajectory state, observability, incident reconstruction

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_x3DtOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether hand-designed trajectory risk accumulation improves pre-execution safety over stateless action filtering. Across the evaluated settings, calibration brings the stateful controller up to—but not beyond—the stateless baseline.

Critical strengths
  1. This paper addresses a practical question and reports a useful negative result.
  2. The evaluation includes multiple models, external trace replay, sensitivity analysis, and evidence ablations.
Critical weaknesses
  1. The theoretical observation is elementary, and the architectural components largely follow established security principles, resulting in limited novelty.

  2. After discovering that feature extraction discarded tool identity, the authors do not evaluate whether retaining that information changes the stateful-versus-stateless comparison.

  3. Retrospective calibration and offline interception counts do not establish generalization or end-to-end safety and task utility.

  4. The experimental diversity is limited. Many runs repeat seven authored attack scenarios; key performance differences concentrate in a single scenario.

Questions for authors

After identifying that the feature projection discarded tool identity, did you reevaluate both controllers with this information retained, including benign run_script calls? Does trajectory state provide any additional benefit under the corrected representation? Can the trajectory controller recover the stateless rule as a special case?

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

The paper asks whether a controller that accumulates risk over an agent's tool-call trajectory prevents more harm before execution than a stateless per-action rule. Across 7 attack scenarios run live on many models, an AgentDojo replay, and real coding sessions, the hand-designed decaying-score state never beats the stateless rule; at the calibrated maximum both block identical runs (Table 4). An elementary collision lemma bounds pre-dispatch policies, and a blinded ablation shows that removing returned tool content yields 33/40 wrong injection-origin attributions.

Critical strengths
  • The negative results are valuable. Tests the common intuition that trajectory state improves agent safety, and honestly reports that it does not.
  • The study covers live attacks on many models, a standard benchmark, and real coding sessions, with sound statistical practice.
  • The diagnostics give concrete guidance for designing tool schemas.
Critical weaknesses
  • The technical contribution for ICLR seems limited. Lemma 1 is self-described as elementary; "Trajectory Control" is not evaluated as a contribution; and the reconstruction result is near-tautological, since origin cannot be recovered once the injecting content is removed. Whether evaluators could abstain is unstated, yet the "confabulate rather than abstain" claim depends on it.

  • The benchmark seems not be able to discriminate the hypothesis. All trajectory discordance comes from one scenario, declaring destruction. The only history-oriented scenario (staged exfiltration) is fully blocked by the stateless rule, and both held-out scenarios yield zero harm. With no live case where history is needed and per-action rules fail, "no advantage" is expected by construction.

  • Although the negative result is valuable, it is quite narrow. Both controllers share a projection that drops tool identity, and the trajectory controller seems impractical on real sessions. The conclusion concerns one linear accumulator over lossy features, not trajectory state; wheras learned and LLM-based sequence monitors are excluded.

Questions for authors
  • Is there any live scenario in which the stateless rule fails but trajectory state would succeed? If not, what result could have supported the hypothesis?
  • How were the held-out scenarios chosen? Since both yield zero harm, what do they test?
  • Why does the shared projection drop tool identity? Was this a design choice?
  • In the reconstruction ablation, could evaluators abstain? If not, how is "confabulate rather than abstain" supported (Table 5)?
Reviewer_xzhkOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper talks about whether a safety guard for AI agents stops more attacks if it remembers earlier steps instead of judging each action on its own. With experiments on a handful of manually crafted samples, the authors say it cannot. However, it is a claim that cannot be directly supported with only seven heuristic scenarios.

Critical strengths

None

Critical weaknesses

1, The content is extremely difficult to understand for any human readers. For example, I still could not even understand the second sentence in Abstract:

"Across separately analyzed live model studies and released AgentDojo traces, the evaluated hand-designed state establishes no prevention advantage over a stateless per-action rule. "

Please at least take 30 minutes to manually write an abstract to be respectful to the reviewers who are trying to understand your work.

Also, in A.16 there are four hashes released, but no actual files to check. It is a very unusual behavior.

2, Seven manually crafted cases could not establish the claim that the authors are making about memory-based guardrails. In fact, there are memory-based defenses in recent published works. [1] https://iclr.cc/virtual/2026/poster/10010390 [2] https://arxiv.org/abs/2605.03228

Questions for authors

1, Could you please rewrite the abstract or reiterate the central idea of this paper? 2, What is the evidence that directly support the this paper's central claim?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.