acceptodds
Under review as a conference paper at ICLR 2027

Available, Relevant, Authorized? Evaluating LLM Agents with Action-to-Evidence

Abstract

Evaluating LLM agents requires distinguishing what they propose, what policy permits, what executes, and what changes outside the model. Conflating these stages can turn missing evidence into apparent safety or credit recovery after an unauthorized effect. We introduce Action-to-Evidence (A2E), a claim-admission mechanism that determines which conclusions execution records support. A2E binds action, policy, invocation, and effect records to the same execution attempt. Effect claims require consistent witnesses; absence and compliant recovery additionally require complete trajectory coverage and observation closure. Incomplete evidence leaves a consistent claim unresolved; conflicting records are rejected as support for that claim. Across 10,934 controlled conformance queries, A2E matches every preset verdict, while removing individual checks admits 801–2,932 unsupported claims. MCP-COMPOSE, our Model Context Protocol benchmark, spans twelve model endpoints, with policy-linked evaluation on five. Replaying frozen proposals isolates the effects of host decisions on execution. In a separate prescribed-action panel with per-call service resets, three invocation-time controls each record 0/180 denied effects but 174–175/180 compliant recoveries. One denied non-target effect followed by a fallback exposes errors in both target-only and fallback-only scoring. On 120 matched attacked AgentDojo trajectories per control, tool filtering and transcribed Progent rules record no benchmark-defined attack effects but complete 2 and 87 tasks, respectively. Evaluations must report effect prevention, compliant recovery, and task utility separately.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, agent evaluation, tool-use security, authorization, prompt injection, runtime monitoring, Model Context Protocol

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

The paper argues that agent evaluations conflate four stages—what the model proposes, what policy authorizes, what the host invokes, and what the service records—so a low count of harmful effects can be misread as safety. It proposes A2E, which binds action, policy, invocation, and effect records to the same execution attempt and checks each claim against claim-specific evidence, returning admitted / unresolved / rejected. Effect claims require consistent witnesses; absence and compliant-recovery claims additionally require complete trajectory coverage and observation closure. The authors build MCP-COMPOSE (12 endpoints, 5 policy-linked) and use AgentDojo to show that equal attack-effect counts can conceal large utility differences.

Critical strengths

• The problem is well identified. The proposal/authorization/invocation/effect distinction is simple but important, and it explains why existing metrics can reach the wrong conclusion.

• The contract is clean and reusable. Same-attempt binding, claim-specific witnesses, and closure-dependent absence/recovery are a genuine conceptual contribution; the action-local vs. trajectory-wide distinction is especially useful.

• The empirical illustrations are compelling. The attrition from 161 tool-name matches to 100 exact linked effects, the 47/83 complete-contract calls, and the “denied non-target effect followed by a fallback” case make the measurement problem concrete.

• The engineering and reproducibility effort is strong. Frozen records, a 167-check consistency audit, clustered CIs, LOEO ranges, and error-inclusive denominators set a good standard.

Critical weaknesses

• The central validation is self-referential. The 10,934 labels are generated by A2E’s own contract, so “matches every preset verdict” is a unit test, not validation. The comparison in Table 7 also uses A2E-defined labels and adapted baselines. Independent human annotation is needed.

• The gate result is close to definitional. Exact PEP uses the same policy oracle that defines denied effects, so 0/180 is guaranteed if the gate works. Recovery is also heavily scaffolded (exact target, exact fallback, two completions, fallback instructions), so 174–175/180 largely measures instruction-following rather than safety capability.

• RQ3 confounds proposal formation with enforcement. Trajectories are generated separately per control, no per-control proposal distributions are reported, and frozen replay is used only in RQ2. The “control location” claim is therefore not cleanly identified.

• The benchmark is small and synthetic. Twelve cases; 180 trajectories = 5 endpoints × 12 cases × 3 repetitions, so inference rests on 12 clusters. The MCP panels are policy-compliance tasks, not prompt-injection attacks, and the AgentDojo attacks are not scored with A2E. The adaptive adversary, alias, retry, and multi-call behaviors named in the threat model are not actually stress-tested.

• The AgentDojo comparison is a strawman for tool filtering. Tool filtering is implemented as a text-only selection request that cripples the agent (2/120); the 70.8/60.0-point gaps are largely artifacts of that choice. Progent is only a hand transcription and its 87/120 is close to Oracle PEP’s 90/120, so the abstract’s juxtaposition is misleading.

• The most important scope limitation: A2E only covers actions that pass through the observed execution boundary. Pure output harms (phishing text, malware code), client-side rendering/exfiltration, bypass execution (code interpreter, direct API calls), delayed or out-of-band effects, human-as-executor, semantic harms, and a bad owner policy all lie outside the evidence domain. The paper excludes some of these in the threat model, but “compliant recovery” and the abstract may lead readers to treat A2E as a general safety guarantee. What it establishes is “no denied effect was observed within the closed window,” not “no harm occurred.”

Questions for authors

See weaknesses

Reviewer_B2FfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Action-to-Evidence (A2E) checks an agent's claims against its execution records. A2E gives each claim one of three verdicts: admitted, unresolved or rejected. A claim about an effect needs matching records from one call attempt. A claim of no forbidden effect needs more. It needs complete records of a finished, fully observed run. Compliant recovery has two parts. The run has no forbidden effect. The run also performs the required fallback. The experiments use 10,934 test cases. They also use MCP-COMPOSE, the authors' benchmark with twelve synthetic tasks. AgentDojo runs compare attack effects with task completion.

Critical strengths

The paper separates five steps: proposal, policy decision, execution, effect and compliant recovery. Two results support this separation. Without a gate, all 56 forbidden proposals execute. The Exact PEP gate blocks all 56. Three adapted judges get the same policy and records as A2E. These judges admit 43 to 95 of the 279 recovery claims labeled false.

Critical weaknesses

On the recovery runs (RQ3), A2E rarely disagrees with simpler scoring. Only one claim differs. In task MCP_DELETE_01, scoring only the target or only the fallback admits one claim of compliant recovery. A2E rejects this claim. A plain count of forbidden effects also rejects it (Table 2). Eleven other runs have incomplete records. A2E leaves their claims unresolved. The MCP setup resets the service before each call. Only sixteen scripted tests cover state that persists across calls. The paper excludes late effects and effects through other channels. The abstract uses missing records to argue for A2E. In this setup, records are almost never missing.

The abstract first reports the 10,934 test cases. Section 4.2 adds 99.07% agreement. Both numbers use labels from A2E's own rules. Section 4.2 states this. With one check turned off, A2E admits exactly the test cases for that check. For example, the full-observation check has 801 test cases. No experiment tests A2E against independent labels.

Questions for authors

The authors should test A2E in one setting with naturally incomplete records. One example is a service with late effects. The paper should then compare A2E's verdicts with effects observed after the run.

Three sets of runs share the name Linked-5. They are the 120-run MCP bridge, the recovery runs (Figure 5, Table 10) and one AgentDojo set. The paper should give each set its own name.

Reviewer_3tMXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Action-to-Evidence, an evaluation framework that determines whether recorded evidence supports a particular claim about an agent’s action. Evidence is bound to the relevant attempt and authorization context, and claims receive admitted, unresolved, or rejected outcomes depending on consistency and completeness. The framework distinguishes positive evidence of an effect from the stronger evidence required to establish absence or successful recovery. Evaluation combines controlled tool-use scenarios, authorization-enforcement comparisons, constructed evidence perturbations, comparisons with adapted evaluators, and an AgentDojo transfer study. The experiments illustrate how apparent task success can conceal an unauthorized effect and how enforcement can prevent that effect while preserving authorized recovery. The contribution is a precise evidence contract and an operational demonstration of its value for agent evaluation.

Critical strengths
  1. The evidence contract captures distinctions that ordinary success labels can miss. Binding evidence to the relevant attempt and authorization context prevents unrelated successes, retries, or fallback actions from being treated as support for the original claim. The separate treatment of positive-effect evidence and complete evidence for absence or recovery is particularly valuable.

  2. The enforcement experiments expose a concrete operational trade-off. In the main denied-action condition, exact enforcement prevents the target effect in all 180 evaluated episodes while preserving recovery in 175. Comparisons with prompting and broader controls demonstrate why a model’s stated compliance and the system’s actual effects must be evaluated separately. The external experiments further examine utility differences among enforcement approaches.

Critical weaknesses
  1. The evaluation establishes contract conformance more strongly than the broader incremental value of the framework. The 10,934 conformance queries include controlled perturbations of evidence records rather than that many independent agent trajectories. These tests are appropriate for checking the implementation, but cannot by themselves establish the frequency or diversity of naturally occurring evidence failures. Comparisons with adapted evaluators are useful, yet the evidence remains limited on how much A2E improves practical evaluation over a simpler provenance-aware deterministic evaluator given the same trusted records.

  2. External validation does not yet cover the complete evidence-admission framework. The AgentDojo experiments primarily evaluate enforcement and task outcomes, rather than the full claim-admission pipeline. Much of the direct A2E validation therefore remains within a controlled scenario family. The persistent-state tests broaden coverage, but do not establish performance when heterogeneous systems produce incomplete, delayed, or difficult-to-bind evidence. This limits the demonstrated generality of the central evaluation contribution, without contradicting the paper’s stated trust model.

Questions for authors
  1. Clarification: Among the unmodified execution records, how frequently does A2E change the conclusion reached from ordinary task-success and policy labels? Please separate naturally occurring cases from deliberately constructed evidence perturbations.

  2. Clarification: Which components of the complete evidence-admission procedure have been exercised on external trajectories? What fraction of claims become unresolved because the required evidence is unavailable, and how does this affect evaluation coverage?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.