acceptodds
Under review as a conference paper at ICLR 2027

Speculative Directive Search for Indirect Prompt Injection Under Multi-Layer Defenses

Abstract

To safeguard agents from indirect prompt injection (IPI), applications layer input detection, prompt enhancement, and model alignment into multi-layer defense stacks. Attacking such stacks raises two challenges: (i) Conflicting constraints: defense layers impose competing requirements on payloads, as explicitness that aids execution also triggers detection. (ii) Sparse feedback: binary success signals under predominantly failing candidates offer little guidance for black-box search. We formulate IPI under multi-layer defenses and propose Speculative Directive Search (SDS). The method adapts the draft-and-verify principle of speculative decoding to black-box attack search. SDS represents payloads through four defense-aligned components. This representation exposes the competing constraints as a jointly searchable space while preserving the target action. In the draft stage, a proxy calibrated on held-out tasks ranks candidates using local detection signals and structural execution features. In the verification stage, the defended agent evaluates candidates in this fixed order under the full stack. Successes undergo repeated evaluation before acceptance. The proxy guides query allocation, while target outcomes determine acceptance. We evaluate SDS across four defense compositions on InjecAgent and AgentDojo. On InjecAgent, SDS achieves an average attack success rate of 72%, compared with 26% for the strongest baseline. These results demonstrate that the tested multi-layer defenses remain vulnerable to structured attack search.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Indirect Prompt Injection, Multi-Layer Defenses, LLM Agents

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

This paper proposes Speculative Directive Search (SDS), a black-box indirect prompt injection (IPI) method targeting LLM agents protected by multi-layer defenses. SDS decomposes attack payloads into four searchable components and employs a calibrated proxy to rank candidate payloads, followed by repeated verification against the target agent.

Critical strengths
  1. The paper investigates prompt injection attacks against combined defenses involving input detection, prompt enhancement, and model alignment. It highlights the potential trade-off between detection evasion and malicious instruction execution, providing a meaningful setting beyond single-layer defense evaluations.

  2. SDS decomposes attack payloads into four components and combines proxy-guided ranking with target-side verification to reduce ineffective search. This structured formulation also facilitates the analysis of different attack factors.

  3. The evaluation covers multiple victim models, defense configurations, and two benchmarks. Component ablations, calibration sensitivity, query-budget analysis, and repeated verification experiments provide useful evidence for understanding the proposed method.

Critical weaknesses
  1. SDS excludes proxy calibration queries from its online budget, while final Beta-CI confirmation requires an additional 6–16 evaluations. Moreover, Table 4 reports an average of 36.4 queries for TAP, exceeding SDS's online budget of 30. It remains unclear whether the total query costs and budget constraints are fully comparable across methods.

  2. SDS trains a gradient-boosted calibration model using success labels obtained from the target defense stack and adopts leave-one-task-out calibration. Although the target task is excluded, the degree of independence between calibration and evaluation tasks remains unclear. In particular, the generalization of a proxy calibrated with only 30 samples to unseen task categories or different defense configurations requires further investigation.

  3. Table 6 shows that fixing the SURFACE component to a reference form reduces ASR to zero across all four defense stacks. This suggests that attack effectiveness may depend heavily on specific payload surface forms or structured-content boundary manipulation, rather than the joint optimization of all four components. The current ablations do not fully isolate these contributions.

  4. In Appendix Table 7, adding Beta-CI confirmation increases the reported ASR compared with majority-vote verification alone. For example, ASR on Stack A increases from 0.94 to 1.00. Since Beta-CI is described as a final confirmation step, this difference requires clarification regarding the evaluation protocol or experimental variability. Additionally, the main results lack corresponding uncertainty estimates.

Questions for authors
  1. Could the authors report the complete target-model query costs, including proxy calibration, online search, repeated verification, and final Beta-CI confirmation, and clarify whether all baselines follow consistent budget constraints?

  2. Are the calibration tasks fully independent of the final evaluation set? Could the authors further evaluate the transferability of the calibrated proxy to unseen task categories or different defense configurations?

  3. Table 6 shows that fixing SURFACE eliminates attack success across all four defense stacks. Could the authors further distinguish the contribution of structured boundary manipulation from the other three components and evaluate whether the method remains effective without relying on such boundary operations?

  4. Could the authors explain why adding Beta-CI confirmation increases the reported ASR in Table 7 and provide confidence intervals or repeated-run statistics for the main results to better establish their reliability?

Reviewer_L7YbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SDS, a black-box attack designed to bypass the multi-layer defenses of LLM agents. To overcome conflicting defense constraints and sparse feedback, SDS decomposes the attack payload into four structural components to formulate a joint search space. The method employs a draft-and-verify principle, using a locally calibrated surrogate model to score and rank candidate payloads before submitting them to the target agent for speculative verification. Extensive evaluations demonstrate that SDS achieves a 72% average attack success rate across four defense combinations, significantly outperforming baselines and revealing critical vulnerabilities in current multi-layer defense systems against structured attack search.

Critical strengths
  • The manuscript is well-written and well-structured, featuring a clear logical flow across different sections.
  • The manuscript clearly articulates the limitations of prior works, the specific challenges encountered, and the corresponding design choices proposed to overcome them.
Critical weaknesses
  • The technical novelty and overall contribution of the proposed method appear to be limited.
  • The authors fail to clearly discuss the SDS’s applicable scope and inherent limitations.
Questions for authors
  • Given the limited technical novelty of this manuscript, I place greater emphasis on its empirical attack performance. Commercial chatbots are widely deployed with significant real-world impact, and it is reasonable to assume that most of them implement multi-layer defenses. I am curious whether SDS remains effective in such practical scenarios. Could the authors evaluate the attack performance of SDS on real-world commercial chatbots (e.g., ChatGPT, Gemini, Claude)? I will consider raising my rating if sufficient results and discussions on this aspect are provided.
  • What is the exact size of the candidate pool used in the final experiments? How does the initialization of this pool affect the overall attack performance? Furthermore, are the results presented in Figure 6 highly dependent on the candidate pool size?
  • Achieving the claimed attack effectiveness requires additional calibration overhead for SDS. However, the authors overlook this in Table 4 and its corresponding analysis. Consequently, the statement "SDS uses 4.4 queries on average..." is somewhat misleading. It is necessary to clarify the total query cost, including the calibration phase.
  • I suggest the authors provide the source code (or at least a core subset) via an anonymous link to ensure reproducibility.
Reviewer_dbLFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Speculative Directive Search (SDS) for indirect prompt injection against combinations of input detection, prompt enhancement, and model alignment. SDS constructs candidates from four components—Surface, Rationale, Directiveness, and Action—and ranks them using a proxy calibrated on separate tasks. Candidates are evaluated in a fixed order against the defended agent, with repeated verification and independent confirmation of returned payloads. The paper reports substantially higher attack success rates than the evaluated baselines on InjecAgent and AgentDojo, together with additional victim-model, SecAlign, and component-ablation results.

Critical strengths
  1. The method addresses a meaningful practical difficulty: an injection must both evade detection and induce the intended action. The structured candidate space and inexpensive ranking model provide an interpretable approach to this joint objective without training an attack-generation policy.

  2. The reported empirical gains are substantial and extend beyond a single victim or single-step setting. The AgentDojo and SecAlign evaluations strengthen the potential practical relevance. The explicit separation between search-side selection and fresh-sample confirmation is also a valuable evaluation design choice, provided its implementation and reporting are clarified.

Critical weaknesses
  1. The reported success and query-efficiency results are not consistently reconciled with the stated protocol. Algorithm 1 requires at least four target evaluations to accept a payload under the default N=3, whereas Table 5 reports median query counts as low as 1.5–2. Figure 7 reports mean query counts above the default budget of 30; the same values appear as query-to-break statistics in Appendix C. Moreover, Table 7 reports higher task-level ASR after adding independent Beta-CI confirmation to N=3 verification, including 0.85 to 1.00 on Stack B. For identical search outputs, an additional confirmation filter cannot increase the number of successful tasks. Different runs or metric definitions could explain these observations, but the manuscript does not establish that explanation. This uncertainty directly affects the central effectiveness and efficiency claims.

  2. The evaluation does not sufficiently isolate the benefit of learned ranking from the candidate library and its handling of input boundaries. Fixing Surface eliminates success across all four stacks in Table 6, while Appendix B describes carrier-specific closing and reopening delimiters. The existing ablations are useful, but the no-proxy ordering in Appendix C is build order, which may be an unnecessarily weak comparator. Random ordering and a best-fixed-template baseline selected on calibration tasks, using identical candidates and verification budgets, would more directly establish the ranking contribution. Clear documentation of the main evaluation's serialization and message roles is also necessary to determine how much the results depend on the particular agent interface.

Questions for authors
  1. In Table 7, do the N=3 and N=3 + Beta-CI rows use the same tasks and search-returned payloads? What explains the increase in task-level ASR after adding confirmation? Please specify the Beta-CI prior, confidence level, stopping rule, and treatment of inconclusive outcomes.

  2. Can you reconcile the query definitions in Tables 4–5, Figure 7, and Table 10, distinguishing first-success queries, verification queries, and costs averaged over all tasks versus successful tasks? Under one common online evaluation budget, does SDS retain its advantage over random ordering and a calibration-selected fixed template using the same candidate library? Calibration costs can be reported separately with their reuse assumptions.

  3. What serialization, message roles, and Spotlighting implementation are used in the main experiments? Can a targeted control that preserves the untrusted tool-data boundary quantify the contribution of delimiter handling? I recognize that the user-role Gemini example in Appendix D is explicitly excluded from the main results; the question concerns the main evaluation configuration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.