acceptodds
Under review as a conference paper at ICLR 2027

Every Guardrail Has a Floor: Deployment Limits of Action Guardrails from the Summary They Read

Abstract

An agent about to send money or delete a file is checked by an action guardrail. Most guardrails decide from a summary of the call: which tool is being used, what kind of resource it touches, how far the effect reaches. Whether the call is an attack turns on a fact the summary leaves out: whether the user asked for it, or an attacker planted the instruction in a document the agent read. In the traffic we study the same bank transfer appears twice, once requested and once injected, with identical summaries. Such a collision has a price that can be computed in advance. From the summary a guardrail reads and a labelled sample of deployment traffic, without its corpus and without its weights, we derive a ceiling on the area under the receiver operating characteristic curve (AUROC) it can reach, a floor on the share of attacks it has to allow at a given retention of ordinary traffic, and the classes of action it has to allow every time. None mentions a model, so they bind every guardrail on that summary at once, no training moves them, and an attacker can read the third off the deployment without querying it. On AgentDojo actions, the tool-name allowlist agent platforms commonly run has to allow of attacks that already succeeded, at retention of ordinary traffic; a static runtime rule . A guardrail certified at recall on a benchmark reaches an AUROC of here, of which retraining recovers only the part its corpus caused. The rest is one field: supplying where the instruction came from moves four language-model judges from an average AUROC of – to –, and a three-valued code carries as much of it as the user's entire task description.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, action guardrails, AI safety, tool-use security, prompt injection, safety evaluation, deployment coverage, provenance

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_mHr6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper identifies a structural limitation of action guardrails. When a guardrail makes a decision from a fixed summary of a tool call, it cannot distinguish an action requested by the user from the same action induced by a prompt injection. This limitation arises because the summary does not contain the source of the instruction.

The work establishes model-agnostic limits that can be computed from the summary and a labelled traffic sample alone. These limits include an AUROC ceiling, a floor on attack pass-through at a given benign retention, and the action classes that every such guardrail must allow. A tool is provided to compute these limits. The work also diagnoses how counterfactual certification benchmarks can write labels into provenance fields, which leaves a near-perfect certificate uninformative about deployment.

The validation includes several components. The authors audit their own record-only guardrail and approximations of tool allowlists, static rules, and permission dialogues on the AgentDojo benchmark. The gap comes down to this one missing field: giving LLM judges the instruction's provenance.

Critical strengths

Firstly, the paper introduces a method to analyze why the guardrails fail during deployment and whether the failure lies in what the guardrail sees or in what it has learned. Using only the fields available to a guardrail and a labelled sample, without access to model weights or the training corpus, a deployer can bound the performance of any guardrail operating on that input and identify the actions that it must allow. Retraining on deployment labels then isolates the part of the gap that is caused by the training corpus. Section 4 further separates boundaries that the input cannot express from boundaries that the corpus mislabels. Although the tools are classical, the diagnostic tells a deployer only the result, not the solution.

Secondly, the paper also exposes a flaw in how agent-safety certification sets are built. Guardrails only learn the provenance fields from certification sets, since the label then becomes a function of them. Removing those fields causes the certified performance to disappear. The coverage analysis in Remarks 1 and 2 then explains why even a near-perfect score provides no information about deployment records that the certification set never encountered. This yields cheap checks that apply beyond this paper, which is another strength.

Critical weaknesses

My main concern is that the authors apply the full certificate-to-deployment audit only to their own guardrail, never to one built by others; the external guardrails (Granite Guardian, Qwen3Guard) are evaluated only through their interfaces on boundaries they never claimed. My suggestion is to test on external papers like ToolSafe or TRIAD. As a result, the paper does not give direct evidence for its claims about real guardrails.

Another concern is that the paper's remedy, an attested provenance field, is the core idea of CaMeL, which the paper cites but does not evaluate. The paper doesn't build such a mechanism itself, and it only tests its provenance proxy against attacks that don't know the defense exists. As a result, the paper shows how much provenance would help, but not how to obtain it reliably, which limits its contribution.

Questions for authors

ToolSafe appears to release TS-Guard with its training data; can you try not auditing it end to end?

Which external, ML-certified, call-only guardrail do the Section 3 certificate results apply to?

Could you report the ceiling and floor when CaMeL's attested provenance is used as the record?

Does the provenance proxy survive an adaptive attack that reformats or encodes injected values?

Reviewer_r8EZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper derives three metrics characterizing the fundamental limitations of action safety guardrails for LLM agents: AUROC, the floor on the share of attacks, and classes of actions. The authors argue that these limitations cannot be overcome by any scoring model or training procedure, and establish achievable upper bounds and lower bounds on the pass-through rates of commonly used guardrails. The authors evaluate these quantities using 2,313 valid actions from AgentDojo and an independent confirmation-dialogue study. By examining three record conditions, Oracle, Provenance, and Neutral, the paper aims to distinguish limitations arising from missing information from those attributable to the scoring model and its training data.

Critical strengths

(1) Clear distinction between representation and model limitations. The paper rightly emphasizes that retraining cannot recover information missing from the input representation, providing a practical record-level approach to diagnose this issue.

(2) Transparent empirical setup. The study is conducted with high transparency, evaluating alternative record mappings, conducting sensitivity analyses, and explicitly reporting a failed preregistered gate in the confirmation study. The accompanying implementation is also a useful practical contribution.

Critical weaknesses

(1) Incremental theoretical novelty and limited generalization. The core results rely largely on standard discrete optimization and statistics over finite partitions defined by the schema. While practically useful, the paper lacks substantive new learning principles or statistical generalization guarantees.

(2) Limited external validity in empirical evaluation. The benchmarks (RecordGuard and MSB-CF) are self-constructed. The benign samples in MSB-CF are created via simple ingress field rewriting, introducing strong artificial shortcuts that may not reflect real-world, independently deployed guardrails.The empirical findings are quite sensitive to labeling rules. Additionally, the headline claim regarding tool-name allowlists uses a randomized LP floor of 40.4%, whereas Table 1 reports a deterministic floor of 86.2% at the same operating point—a major difference that needs explicit clarification.

(3) The typesetting and document layout deviated from the official latex template of ICLR, making the paper look unprofessional and difficult to read.

Questions for authors

(1) Novelty of Theoretical Contributions relative to Prior Work. The core result of Proposition 1 essentially boils down to constrained discrete optimization over finite record cells. The paper should more clearly articulate what novel theoretical guarantees this framework offers compared to existing literature on constrained classification and optimal decision analysis, beyond merely applying established optimization formulations to guardrail evaluation.

(2) Reliability of Empirical Lower Bounds in Real-World Deployments. Appendix F indicates that varying sample weights and label-matching rules noticeably impact the estimated pass-through lower bound. I invite the authors to clarify how stable these bounds remain under independently sampled deployment traffic, and whether formal statistical uncertainty analysis can be provided. Should the current results be interpreted primarily as an audit tailored to the specific empirical distribution of AgentDojo?

(3) Validation on Independently Developed Guardrail Systems.While the authors present supplementary analyses using ASB-lite and other guardrail interfaces, the most comprehensive certification audit relies on RecordGuard, which was developed by the authors themselves. Validating the framework on independently developed guardrails using their native certification data would help determine whether the identified vulnerabilities are equally prominent in real-world deployed systems. The authors should further clarify the explicit boundary and scope of their current empirical conclusions.

Reviewer_Ej5ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the limitations of summary-based action guardrails and derives model-independent bounds on their AUROC ceiling, harmful pass-through floor, and must-pass action classes. Experiments on AgentDojo and τ²-bench demonstrate these limitations and show that adding instruction provenance or relevant context can significantly improve guardrail performance.

Critical strengths
  1. Sound theoretical analysis. The paper provides a useful framework for quantifying the performance limits of summary-based guardrails, independent of specific models or training data. And the performance on benchmark cannot stands for the effect in real-world deployment.

  2. Useful empirical findings. The experiments demonstrate how missing context and provenance can limit guardrail performance, and how providing this information can improve detection.

Critical weaknesses
  1. Limited practical relevance. The main findings rely on restricted action summaries that make benign and malicious actions indistinguishable. However, modern guardrails, such as those in Codex and Claude Code, can use user requests and execution context. It is unclear how well the reported limits generalize to these systems.

  2. Lack of evaluation on realistic guardrails. The paper mainly evaluates a self-developed record-based guardrail, without sufficient comparison to modern context-aware or trajectory-aware guardrails. This makes it difficult to assess the practical significance of the proposed bounds.

Questions for authors
  1. How representative are the summary-based guardrails studied in this paper of modern deployed agent systems? For example, systems such as Claude Code and Codex can use the original user request and relevant execution context when checking tool calls. Could the authors clarify which real-world guardrail implementations rely on input representations similar to those evaluated in the paper?

  2. Table 3 shows that providing additional context or provenance substantially improves detection performance. Given that the reported pass-through floor drops to zero when full tool arguments are available, to what extent do meaningful non-zero floors remain for context-aware guardrails?

  3. The paper evaluates Granite Guardian and Qwen3Guard, but these models are primarily designed for general safety moderation rather than tool authorization. Could the authors include experiments with more representative agent guardrails that explicitly check tool calls against user intent and execution context, to assess whether the proposed limitations persist?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.