acceptodds
Withdrawn from ICLR 2027

PACE: Provenance-Aware Capability Enforcement for Tool-Using LLM Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Tool-using large language model (LLM) agents turn generated text into real side effects, so poisoned tool metadata, retrieved pages, memory, and reusable skills can steer the next call. Vetting an artifact before admission does not settle this. A safe variant and a leaking variant can produce the same admission evidence, and a sound gate then cannot relax that site for either. We make that condition precise, which leaves the last boundary a deployment can still act on. We present Provenance-Aware Capability Enforcement (**PACE**), which mediates every tool call immediately before it executes. Path confinement proposes an executable cut of represented influence paths, while capability and effect verification checks schema-defined effects against authority compiled from the authenticated request. We distinguish the certified execution contract from the evaluated configuration, which can restore an authorized call after a proposed block or apply a declared repair. Confinement requires the final action to preserve the certified cut. On eight executable agent-security benchmarks with three target-model families, the evaluated configuration gives strictly lowest attack success in **62 of 79 eligible attack columns** and ties in 14; full-benchmark native utility loses at most three points relative to the undefended agent. A complete ablation over 1167 paired cases attributes most security gains to effect verification and refusal control to boundary adaptation. A reduced-scale adaptive search succeeds on 0/30 out-of-authority targets against the defense. Source code is in the supplemental material.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Agent Attack Defense, LLM Agent Safety

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_9CrvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PACE, a runtime defense for LLM agents that combines path confinement, capability and effect verification against authenticated user requests, and execution boundary adaptation. It analyzes the limits of admission checks and provides conditional guarantees for represented paths and effects, while distinguishing its certified contract from the evaluated configuration and avoiding claims of full noninterference. Across eight benchmarks and three model families, PACE achieves the lowest attack success rate in 62 of 79 eligible comparisons and ties in 14. The evaluation also includes component ablations on 1167 paired cases and adaptive attacks against 30 targets screened as outside the compiled authority.

Critical strengths

1.This paper clearly separates path confinement from effect authorization and explains why restoring an authorized call may invalidate a path guarantee. Its explicit treatment of assumptions and uncovered channels makes the formal analysis useful for understanding the defense’s scope. Although I'm not fully clear about some of the details, I think the overall writing of the paper is good.

2.The evaluation spans multiple benchmarks and model families while measuring both security and utility. The complete component ablation and disclosure of substantial utility losses on some subsets support a more balanced assessment of the system.

Critical weaknesses

Given the paper’s substantial reliance on its implementation, I examined the accompanying code and identified several issues that raise concerns about whether it adequately supports the paper’s claims.

  1. The supplied implementation does not adequately establish the semantic assumptions underlying its guarantees. In offline checks, requests prohibiting email or restricting delivery to Alice still allowed the effect checker to certify an email addressed to Mallory when the adapter did not flag attacker influence. Explicit structured recipient constraints correctly blocked the call. Additionally, the get_webpage tool is classified as reversible with no sink, allowing a tainted URL through R2 despite a failed CTRL check. For actual HTTP execution, this overlooks disclosure through request URLs. These findings weaken confidence that the evaluated compiler and schemas faithfully represent user authority and concrete effects.
  2. The connection between the proposed path mechanism and the evaluated implementation remains unclear. The default one_hop_fallback path construction contains no committed historical edges and supports only the block_call action. This remained true after a synthetic tool result was committed, and I could not locate an adapter supplying complete historical snapshots in the provided code. Several simpler ablation variants match the full system’s attack success rates, while paired AgentDyn utility falls from 66.7% to 33.3%. Missing frozen manifests and episode records further prevent independent verification of the implementation used and the contribution of path confinement.
Questions for authors
  1. How does the evaluated compiler enforce negation, recipient restrictions, and resource constraints? Which layer accounts for data transmitted through URLs and query parameters?
  2. What proportion of evaluated calls used complete historical snapshots versus one_hop_fallback? Please provide an actual trajectory connecting prior provenance, the certified cut, and the executed intervention.
  3. Can the authors provide the frozen manifests, episode records, table regeneration scripts, and corresponding code revision? Were the documented MELON embedding and DataSentinel backbone substitutions used in the reported experiments, and how was their fidelity validated?
  4. Among the 254 pairs that passed the AUTH prescreen, how many fall within the compiled capability envelope while still violating the user’s intended resource or argument restrictions?
Reviewer_i4vWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies prompt injection and unauthorized execution in tool-using LLM agents and introduces PACE, a runtime defense framework. Before a tool call is executed, PACE combines path constraints over a provenance graph with authorization and effect verification grounded in the trusted user request to determine whether to allow, block, or modify the call. Experiments across multiple security benchmarks and models show that PACE reduces attack success rates while maintaining good utility on benign tasks.

Critical strengths
  1. As LLM agents gain permissions to access files, send emails, and perform other operations, prompt injection can lead to sensitive information disclosure and unauthorized external actions. Mitigating these risks while preserving task completion capabilities is an important challenge for deploying agents in practice.

  2. The paper compares PACE with several categories of representative defenses across eight benchmarks and three target model families. By evaluating both attack success rates and task utility, the experiments provide a comprehensive assessment of its defensive effectiveness and its impact on benign tasks.

Critical weaknesses
  1. Limited analysis of individual intervention types. PACE supports several interventions, including blocking, argument rewriting, read-only views, and budget restrictions. However, the paper does not sufficiently analyze how frequently each intervention is applied or how effectively it contributes to the overall defense.

  2. Missing analysis of runtime efficiency and overhead. PACE performs provenance tracking, path checks, and effect verification before each tool call, but the paper does not report the resulting latency or resource overhead.

  3. Lack of a direct comparison between PACE-C and PACE-P. The theoretical guarantees primarily apply to PACE-C, whereas all experiments evaluate PACE-P. The latter allows calls blocked by path checks to be reinstated, and some repairs are applied without revalidating the final effects. A direct comparison of the two configurations in terms of security and task utility would clarify the practical impact of these adjustments.

  4. Insufficient fine-grained ablations of the provenance graph and path-cutting mechanism. The provenance graph tracks data sources and dependencies, while the path-cutting mechanism determines where and how to intervene. Both directly affect defensive effectiveness and utility on benign tasks. However, the paper lacks separate ablations of provenance graph construction and path-cutting strategies, making it difficult to assess the contribution of each design choice to the final results.

Questions for authors

See weakness above.

Reviewer_pe2zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces PACE, a provenance-aware security framework designed to defend tool-using LLM agents against prompt injection attacks. Rather than relying solely on detecting malicious instructions, PACE enforces security at the tool-execution boundary by combining provenance-based path confinement with capability and effect verification. The authors provide formal security guarantees under specific assumptions and evaluate the framework across eight benchmarks and three model families. The results show substantial reductions in attack success rates while largely preserving the agents' utility.

Critical strengths
  1. Strong Defense Performance: PACE achieves impressive results, significantly reducing attack success rates across eight benchmarks and three model families while maintaining utility close to the undefended agents.

  2. Interesting Defense Approach: Enforcing security at the tool-execution boundary rather than relying solely on detecting malicious instructions is an interesting idea. The combination of provenance tracking and capability-based verification seems promising for securing tool-using agents.

  3. Comprehensive Evaluation: The evaluation is quite extensive, covering diverse benchmarks and existing defenses. I also appreciate the inclusion of ablation studies and adaptive attack experiments, which provide useful insights into the effectiveness of the proposed approach.

Critical weaknesses

Major Concerns:

  1. Security Correctness: PACE heavily relies on a trusted registry to understand tool behavior and determine the consequences of tool calls. I think a critical weakness here is the unrealistic assumption that these descriptions are always available, accurate, and up to date. This is particularly concerning in dynamic tool ecosystems like MCP, where tools are frequently added or updated. Incomplete or outdated descriptions could easily undermine PACE's security guarantees.

  2. Runtime Overhead: While PACE's defense results look promising, I would really like to see the authors report the runtime overhead in terms of latency, token usage, additional calls, etc. I think these numbers are important for understanding how practical PACE is and making a fair comparison with existing baselines.

Questions for authors

N/A

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

Related papers

Discussion (0)

Sign in to comment.