acceptodds
Under review as a conference paper at ICLR 2027

Beyond Prompt Filtering: What Runtime Safety Mechanisms for LLM Agents Observe, Decide, and Control

Abstract

Runtime defenses for tool-using LLM agents operate at different observation, decision and execution interfaces, making their safety claims difficult to compare. We present a source-anchored comparative synthesis of 77 version-pinned comparison records spanning runtime mechanisms, evaluations and related syntheses. A Horizon–Modality (H/M) map characterizes the evidence available to named checking stages and how they decide across 25 core mechanisms and supporting components. Mechanisms grouped under information-flow defenses can rely on learned behavioral judgments or enforced runtime restrictions, requiring different validation evidence. The synthesis distinguishes task alignment, authorization and effect correctness, and identifies validation obligations for policy updates and cross-authority handoffs. Two external human coders independently annotated 68 core/extension records, agreeing on all 25 core H/M assignments; agreement was lower for decision targets and inferred premises. We map evaluations by whether they measure judgments, executed effects or legitimate completion, clarifying what evidence supports a learned guard's deployment claim. A comparison-record schema, validator and worked examples support reuse, alongside a handoff checklist, reporting template and an executable handoff test with 80 fixtures and expected outcomes.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, multi-agent systems, runtime enforcement, information-flow control, agent safety

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_d5YYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

My reading is that the paper tries to clarify what runtime defenses for LLM agents can observe, decide and actually enforce. It brings together 77 version-pinned records with 25 core mechanisms mapped in detail. I see the external coding and small handoff tests as support for this comparison.

Critical strengths

I appreciate the care put into the source records. Separating a learned safety judgment from an enforced restriction also helps me make sense of the different systems. The authors show where the coders disagreed, which I found usefull. The paper has value as a reference.

Critical weaknesses
  1. My main concern is that I don't see enough of a new contribution. Section 2 explains how this map differs from earlier surveys but the main takeaways still feel familiar: local checks may not compose, permission doesn't establish intent, and a safety verdict doesn't tell us what actually happened. Organizing these ideas is useful. What I'm missing is a substantial conclusion or design insight that emerges from putting them together. A survey can certainly make that contribution and I don't think the finer categories here are enough on their own.

  2. I'm also not convinced that someone else could use the framework reliably. In Appendix B.1, exact-set agreement is 54.5% for target labels and 47.7% for inferred-premise labels across 44 jointly applicable records. The core horizon, modality, authority-scope labels do agree, which is reassuring. The harder distinctions remain the issue: having an author settle the disagreements doesn't show that another reader would reach the same decisions. Nor do the 80 handoff runs in Section 6.2 close that gap. They check expected outcomes on examples the authors constructed, including deliberately permissive controls. I'd find a case where the framework exposed a previously overlooked problem much more convincing. The authors are upfront about these limits but the evidence is still too limited.

Questions for authors
  1. I may be missing something: which finding goes furthest beyond what the surveys in Section 2 already tell us?

  2. Do you have an example where someone independently used the framework and found an overlooked issue or changed a design decision? I'm asking about existing evidence here rather than proposing another experiment for the rebuttal.

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

This paper reviews runtime safety mechanisms for LLM agents. It compares what these mechanisms observe, how they make decisions, and how they control execution. It also discusses their trust assumptions and evaluation methods, and provides comparison records and a checklist for agent handoffs.

Critical strengths

1 The source-linked comparison records provide a useful reference for future research. 2 Independent human coding and explicit discussion of limitations improve the transparency of the study.

Critical weaknesses
  1. The main contribution is descriptive. The paper organizes existing ideas well, but its new technical insights are relatively limited.
  2. The paper selection is partly based on author judgment. It remains unclear how much the conclusions depend on the selected systems.
  3. Some categories are difficult to apply consistently. The lower agreement on decision targets and inferred assumptions suggests that these categories need clearer definitions.
  4. The practical benefit of the framework is not fully demonstrated. The handoff experiment mainly checks predefined rules and provides limited evidence about improvements to real agent safety.
Questions for authors

NA

Reviewer_26mbOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues that local guarantees provided by existing LLM agent safety mechanisms do not necessarily translate into system-level runtime safety. Different mechanisms vary in their observation scope, decision objectives, and enforcement capabilities. When combined, they may also introduce security vulnerabilities if authorization information, data provenance constraints, or execution states are not properly propagated. To address these issues, the authors propose a taxonomy based on observation scope and decision-making approaches, and further analyze the conditions for combining safety mechanisms and the security claims supported by existing evaluations.

Critical strengths

The paper provides a systematic review of existing runtime safety mechanisms for LLM agents and compares their designs and actual guarantees in terms of observation scope, decision-making approaches, intervention methods, and trust assumptions.

Critical weaknesses
  1. The paper introduces numerous classification dimensions, symbols, and terms, but the relationships among these concepts and their roles in supporting the main conclusions remain unclear. For example, Section 4 analyzes the designs, assumptions, and limitations of existing safety mechanisms individually, but the comparisons have not been sufficiently distilled into broader conclusions. As a result, it is difficult to determine what new insights the proposed taxonomy offers beyond a systematic review of existing work, and how these insights can guide the design and composition of practical safety mechanisms.
  2. The experiments mainly automate predefined logical checks. Since the test policies, data labels, and expected outcomes are specified in advance, it remains unclear what additional value these experiments provide compared with manually checking each case. Furthermore, the paper does not demonstrate whether the experiments can uncover security issues that are difficult to identify through manual analysis.
Questions for authors
  1. Compared with existing taxonomies and surveys of safety mechanisms, what substantive new insights does the proposed taxonomy offer? Can these insights be translated into concrete design principles for individual mechanisms or their composition?
  2. What additional findings does the experimental evaluation provide compared with manually checking each case? Can the authors demonstrate that the proposed framework identifies security issues that are difficult to uncover through manual analysis?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.