acceptodds
Under review as a conference paper at ICLR 2027

After Denial, What Next? Recoverability-Aware Legal Continuation for Tool-Using Agents

Abstract

Tool-using agents pursue user goals through calls to external services. Authorization determines which tool actions are permitted, but a denial leaves an unresolved decision: which permitted action should the agent take next? A poor continuation can stall a recoverable task or request unnecessary capability. Existing authorization checks enforce the legal boundary without ranking the ways forward. We formulate post-denial legal continuation and introduce Denial-Aware Recovery Projection (DARP). Given structured denial feedback, DARP constructs typed recovery candidates, checks them against the current policy, and estimates their bounded-horizon task value. It selects among near-maximal legal candidates by capability exposure, execution cost, and repair distance, then binds the chosen action to a one-time state-bound certificate for dispatch. On 320 template-held-out tasks with registered deterministic transitions, exact values, a three-step horizon, and complete candidate recall, DARP achieves 100% safe resolution with 27% less mean capability exposure than a bounded-horizon legal value filter. Replacing exact values with a frozen Qwen scorer yields 90.94% safe resolution with no unauthorized dispatch.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Tool-Using Agents, Agent Safety, Post-Denial Recovery, Least Privilege

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_T44eOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

1.The authors propose DARP, which turns structured denial feedback into typed alternatives, compares their recoverable value within the legal set, and limits capability use among near-maximal choices. State-bound certification carries the selected action through dispatch before the agent replans from the observed return.

Critical strengths

1.The results are compeling: DARP reduces mean capability exposure from 10.20 to 7.44. Replacing exact values with a frozen Qwen scorer yields 90.94% safe resolution and zero unauthorized dispatch.

2.The studied topic is valuable and offers real-world impact.

3.Metrics are multi-facet and evaluates the phenomenon well.

Critical weaknesses

1.The authors can better explain the terms used in Section3 Formulas, e.g., write the full names of each term in Formula 1. Also, Cap(a) is the semantic capability needed to execute a. So why not define it exe(a)?

2.The Figure 1looks a bit junior, with much important information omitted.

3.The details about benchmark construction are largely omitted, e.g., how many authors participate in the annotation?

4.Some parts of writing is hard to understand. e.g., line 186-187.

5.For the correct stop rate, authors can record agents stop on which steps, which can help better measure token wastes.

Questions for authors

See above.

1 comment
AuthorsOpenReview ↗

Response to Reviewer T44e

Thank you for your review.

Q1. ... Formula 1 ... Cap(a) ...

Answer: The revised Eq. (1) names all six fields: operation, arguments, preconditions, effects, required semantic capabilities, and requested provider scopes. Cap⁡(a)\operatorname{Cap}(a) is the authority an action requires. Exec⁡\operatorname{Exec} runs the action and returns its observation. For a permission request, Req⁡(a)\operatorname{Req}(a) records the scopes being sought.

Q2. ... Figure 1 ... information ...

Answer: The redrawn Figure 1 shows the six stages of the recovery loop and the objects passed between them. It highlights how equal value and lower exposure determine at∗a_t^* in a recorded Gmail decision, then follows that action through verification, dispatch, and replanning. Table 14 gives the legality, value, and exposure of each candidate.

Q3. ... benchmark construction ... authors ... annotation ...

Answer: The revision identifies two authors who constructed or reviewed the templates, gold labels, and evaluation rules in Section 4 and Appendix I (Benchmark construction and example, p. 17). Appendix C specifies that two authors independently annotated each of the same 80 blinded cases, with randomized candidate order and with method identities, selected actions, scores, and evaluator labels withheld.

Q4. ... hard to understand ...

Answer: The revised typed-recovery paragraph after Eq. (9) states the candidate processing rules. The schema validates the arguments, and equivalent candidates are merged by a semantic fingerprint that omits the candidate ID. A non-repeatable action is retained if a modeled outcome changes a fact, adds a capability, or terminates the episode. Safe stop is registered explicitly.

Q5. ... correct stop rate ... steps ...

Answer: Table 23 reports correct stops, the post-denial step, and premature stops on recoverable tasks. Exact-value and Qwen-estimated DARP correctly stop at the first decision on all 40 unrecoverable tasks. Self-Revision and Direct Retry make their correct stops at the same step, on 37 and 38 of the 40 tasks. Appendix I (Stopping behavior, p. 22) reports 775.45 incremental scoring tokens per unrecoverable task for Qwen-estimated DARP with cached proposals, separately from the table. The stopping analysis makes no end-to-end efficiency comparison.

Q6. ... See above ...

Answer: The revised Section 3, Figure 1, benchmark and human-audit protocols, and stopping analysis address Q1–Q5 above. Is there a specific assumption or step in the soundness argument that needs further clarification?

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

This paper introduces a recovery projection system that formalizes what an agent should do when a tool call leads to a denial. In such situations, it is generally unclear which permitted action the agent should take next. Existing work such as Progent or MiniScope can formalize which tools an agent may call, but not what it should do after a tool call is denied. The authors thus introduce DARP, which selects the gold action when available and allows the model to recover safely. Alongside the method, the DARP-Post-Denial benchmark was proposed. With formal guarantees, the authors show that their approach beats existing baselines on DARP-PostDenial. The paper also includes statistical and human audits.

Critical strengths
  • The problem is relevant and the formulation novel. Selecting the best action to take after denial is an interesting research question, since existing methods focus only on denying unauthorized tool calls and do not take the steps to select an alternative that is both safe and aligned with the original task.

  • On three different model families (Qwen, Claude, GPT), the proposed method meaningfully increases the performance of baselines and ablations on the author's proposed benchmark.

Critical weaknesses
  • Evaluation setup: The method is only evaluated on the author's own benchmark, DARP-PostDenial, which was fully co-designed alongside the method. This creates a less fair comparison: it is not clear if the proposed method would work on other benchmarks, for instance standard tool call benchmark where the gold action might not be clear, or indirect prompt injection benchmarks where the prompt injection might interfere with the selected gold action. Including the original evaluations benchmarks tasks used in Progent or TS-Flow would allow for a fairer comparison. Reusing the original Progent and TS-Flow public implementations would also strengthen the evaluation.

  • Writing and missing details: There are multiple key missing details in the paper, such as the prompts used to generate the benchmark, the origins of the Google Workspace documents mentioned in line 710, and more details on the statistics of the proposed benchmark. For clarity, it would also be good to add specific examples of benchmark tasks, and comparisons with other tool call benchmarks. The prose in many sections such as 1 and 3 also feels automated, which makes the paper harder to understand. Revising the AI-assisted text futher would strongly increase the paper's quality.

Questions for authors

Besides the questions raised above, I would like to ask the following questions:

  • Line 289: baselines such as Progent-Fallback and ToolSafe TS-Flow are reimplemented, but these works already have a public GitHub with an official implementation. Which steps were taken to show that these reimplementations respect the original paper, and why were the official implementations not used?

  • Does DARP achieve good performance on independently created benchmarks, and are the proposed recovery tool correct on these benchmarks?

  • The baselines Progent and TS-Flow were specifically designed for defending against indirect prompt injections. What is the performance of DARP on indirect prompt injection benchmarks (for instance on AgentDojo)? Do the formal guarantees hold? In particular, making sure that post-denial recovery on prompt injection does not lead to the gold answer being influenced by the prompt injection would also be important.

2 comments
AuthorsOpenReview ↗

Response to Reviewer nEwf

Thank you for your review.

Q1. ... gold action ... recover safely ...

Answer: DARP selects from the legal set by bounded-horizon value and the capability-based rule. The revision separates the registered planning actions and utility predicates used for scoring from the gold labels used for CR and FAA in Section 3 and Appendix D. Perturbing evaluator labels preserves the public-view, prompt, and cache hashes of all 640 tasks and the hashes of all 480 available candidate artifacts.

Q2. ... co-designed ... original evaluations ...

Answer: The revision adds a comparison of original Progent with Progent plus DARP on independent AgentDojo tasks in Appendix I (Independent environment, p. 20). Both arms share the GPT backbone, upstream policy generator and updater, secured wrapper, and AgentDojo outcome checks. In four induced denials the arms match on the rejected call, state, policy, and visible history. AgentDojo evaluates task utility without a gold next action. Table 22 separately identifies the shared-checker Progent-Fallback and TS-Flow controls on DARP-PostDenial.

Q3. ... prompts ... documents ... statistics ... prose ...

Answer: Appendix I (Benchmark construction and example, p. 17) gives the 30 ID templates, template-disjoint 80/80/320 splits, ten unseen OOD templates, and domain and denial counts. Each template has two fixed resource records with explicit objectives, targets, and payloads, each instantiated under eight denial conditions. The Discovery descriptions supply API methods and scopes; the resource records supply the mail, file, and calendar fields. Appendix J (Evaluation prompts and recorded inputs, p. 22) contains the complete evaluation prompt templates and a recorded Gmail input and output. The anonymous review artifact includes the full request/response pairs under prompts/examples/ and all two-instance records in benchmark/resource_instances.json. Figure 1 and the appendix give Gmail, Drive, and Calendar examples and distinguish the benchmark from API-Bank, ToolSandbox, and AgentDojo. The revised Section 1 and Section 3 explain the recovery decision before its formal notation.

Q4. ... reimplemented ... official implementation ...

Answer: Table 2 and Table 22 identify the ID/OOD baselines as shared-checker adaptations of published recovery mechanisms. The adaptations fix Qwen, tool schemas, the three-step budget, gateway, and executor while varying the recovery policy. Progent-Fallback preserves policy-failure replanning and uses the common checker's failed predicates in place of the native error. TS-Flow preserves feedback-driven retry and uses the shared verdict and predicates in place of TS-Guard. Seventeen review-stage checks cover the tested control branches and adapter contracts, including threshold and retry with fixed guard scores. Appendix J (Evaluation prompts and recorded inputs, p. 22) shows the recorded checker feedback and fallback prompt. Appendix I (Independent environment, p. 20) reports the separate AgentDojo experiment with the original Progent implementation and runtime.

Q5. ... independently created benchmarks ... correct ...

Answer: The new 12-task AgentDojo comparison in Table 21 gives 10/12 benign and 9/12 attacked task completions in each arm. In the four matched, induced denials, both arms execute a legal next call in every case and each completes two tasks. All four DARP replacements pass the original policy and certificate checks before the original runtime executes them. The revised table distinguishes next-call execution from eventual task completion.

Q6. ... prompt injection ... formal guarantees ...

Answer: The new attack comparison in Table 21 gives 9/12 legitimate completions and zero attacker successes for both arms. Both execute a legal next call in the two matched denials under attack. The revision explains in Section 3 and Appendix I (Independent environment, p. 20) that injected tool returns may influence proposals and ranking, while the authenticated objective and original outcome checks remain fixed. Dispatch legality requires sound checking and atomic certified dispatch against trusted state and policy. The separate ϵ+2δ\epsilon+2\delta bound applies within the legal candidate set under uniformly bounded value error.

Thank you for the response and the additional experiments. The new AgentDojo experiment adding DARP to Progent is interesting, but with 12 tasks and only four denials, the outcomes are the same and the experiment is too small to tell whether DARP improves recovery. Together with the presentation, which was improved but still hard to follow in places, and the concerns the other reviewers also raised, I will keep my score.

That being said, I think post-denial continuation is an interesting research problem. I encourage the authors to strengthen the evaluation on independent benchmarks, to clarify the robustness guarantees, improve the presentation, and to resubmit a revised version to a future venue.

Reviewer_7icxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how tool-using agents can continue a task after a tool call is denied while respecting authorization constraints. It proposes DARP, which generates typed recovery actions from structured denial feedback, filters them for legality, estimates their recoverable task value, and selects among near-optimal candidates using a lexicographic preference for lower capability exposure, execution cost, and deviation from the denied action. The experimental section demonstrated their advantages.

Critical strengths

The paper provides a clear and principled perspective by treating denied tool calls not merely as terminal failures, but as decision points that balance authorization constraints with task progress. Decoupling legality checking from recovery planning provides a systematic framework for evaluating compliant fallback actions.

Critical weaknesses
  1. Figure 1 show the pipeline through text-heavy boxes, without illustrating a recovery example showing candidate actions, their estimated values and capability requirements, and how the final action is selected. â_t in figure but a* in method?
  2. The method part is a liitle bit hard to follow, the overall description is not coherent enough.
  3. The ratios in Tables 1 and 3 are rounded to one decimal place (e.g., 100.0, 84.3), while the same ratios in Tables 2, 4, and 5 are rounded to two decimal places (e.g., 100.00, 99.64).
  4. One concern is that DARP’s strongest results rely on exact values in a deterministic sandbox with a three-step planning horizon. Replacing exact values with a frozen Qwen scorer reduces SRR to 90.94%, below Self-Revision’s reported 95.9%. This sensitivity to value estimation leaves DARP’s practical advantage under uncertain tool dynamics insufficiently established.
Questions for authors
  1. ReGuard (Pandya et al., 2025), Agent-C (Kamath et al., 2025), AgentSpec (Wang, Poskitt, and Sun, 2026), these papers are some related works, if it possible to compare these with your method? If not, please give more clarification.
  2. What proportion of the finally selected actions at∗a_t^* come from deterministic control instantiation versus LLM generation? How does DARP perform if deterministic control candidates are ablated and all recovery candidates must be proposed by the LLM?
  3. Formula (9) defines the operator as a mapping ω:(q,st,dt)↦{(a,preω,effω)}\omega: (q, s_t, d_t) \mapsto \{(a, \text{pre}_\omega, \text{eff}_\omega)\}, but in Formula (1), the action aa is already a six-tuple containing pre(a)\text{pre}(a) and eff(a)\text{eff}(a), resulting in overlapping and confusing data structure definitions. Can author give more explaination?
  4. The other questions are in weakness.
2 comments
AuthorsOpenReview ↗

Response to Reviewer 7icx

Thank you for your review.

Q1. ... Figure 1 ... recovery example ...

Answer: The revised Figure 1 shows how the checker, value frontier, exposure comparison, and dispatch verifier connect. In the recorded Gmail decision, the two legal requests have equal value (0.925) and incremental exposures of 19 and 22. The narrower request is selected as at∗a_t^*, matching Eq. (18). The figure shows why retry is rejected and why legal stop falls below the cutoff. Table 14 gives the four candidate records.

Q2. ... method part ... coherent ...

Answer: The revised Section 3 follows one decision from KtK_t through the legal set Kt0K_t^0 and value frontier KtϵK_t^\epsilon. It then shows how the lexicographic key selects at∗a_t^* and how the certificate binds that same action at dispatch. The Gmail case in Figure 1 follows the same order with concrete actions and values.

Q3. ... ratios ... decimal place ...

Answer: The revised result tables use two decimal places for comparable percentage-rate columns. The values in Table 1 and Table 3 are rounded from the original runs and now follow the same convention as Table 2, Table 4, and Table 5.

Q4. ... exact values ... Self-Revision ... SRR ...

Answer: The new paired analysis in Appendix I (Matched estimated-value comparison, p. 18) reports 100% CSR for Qwen-estimated DARP versus 92.50% for Self-Revision, and mean exposure of 8.77 versus 9.47 on 243 shared completions. Full-cohort SRR is 90.94% versus 95.94%. All 29 unfinished recoverable DARP cases end at stop. In 18, the best legal non-stop candidate scores at or below stop; in 11, it scores higher but stop remains in the value frontier. That comparison holds the registered dynamics fixed. The revision separately adds an AgentDojo experiment that scores from the visible interaction history and executes against observed tool returns in Appendix I (Independent environment, p. 20).

Q5. ... ReGuard ... Agent-C ... AgentSpec ...

Answer: The revised Section 2 distinguishes their decision interfaces. ReGuard learns a predictive safety monitor and recovery policy. Agent-C generates actions under SMT-checked temporal constraints. AgentSpec applies rule-triggered interventions. DARP ranks actions already admitted by a checker after an explicit denial. A matched numerical comparison would require a common task, observation, and authorization interface. The revision evaluates that interface with the shared-checker controls and the new upstream Progent experiment in Appendix I (Independent environment, p. 20).

Q6. ... deterministic control instantiation ... LLM ...

Answer: The revision reports the source of all 420 selected recovery actions in Appendix I (Proposal provenance and coverage, p. 19). Of these, 260 (61.90%) are model proposals, 120 (28.57%) are deterministic control instantiations, and 40 (9.52%) are safe stops. In the new proposal-source ablation, removing deterministic proposals at every decision while keeping safe stop and exact values gives 87.50% candidate recall and 98.13% SRR. Exposure rises from 7.44 to 12.81, with zero UDR in both conditions (Table 17).

Q7. ... Formula (9) ... operator ...

Answer: The revised Eq. (9) returns a set of registered typed actions, ω(q,st,dt)⊆Aω\omega(q,s_t,d_t)\subseteq\mathcal A_\omega. Each returned aa already carries pre⁡(a)\operatorname{pre}(a) and eff⁡(a)\operatorname{eff}(a) as defined in Eq. (1). Checking, planning, and certification use that same action representation.

Q8. ... other questions ... weakness ...

Answer: Q1–Q4 address the points raised above. The corresponding revisions are in Figure 1, Section 3, the result tables in Section 4, and the new estimated-value analysis in Appendix I (Matched estimated-value comparison, p. 18).

Thanks your response. Given the overall layout and quality of the current paper, I will maintain my original rating.

Reviewer_MKB4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how tool-using agents can continue a task under authorization constraints after a tool call is denied. The authors propose DARP, which generates and filters legal recovery actions based on denial feedback, evaluates their recovery value through bounded-horizon search, and prioritizes actions with lower capability exposure. The selected action is revalidated before execution.

Critical strengths

The problem is practically relevant. Rather than simply blocking unauthorized tool calls, the paper considers how agents can recover and continue their tasks after denial.

The framework is relatively complete, covering authorization checks, recovery decisions, and execution verification. The experiments also include baseline comparisons and ablation studies.

Critical weaknesses

The main results rely on deterministic state transitions and exact task values. When using Qwen for value estimation, the safe resolution rate drops from 100% to 90.94%. Capability exposure results and comprehensive baseline comparisons under this setting are also missing.

The evaluation is limited to synthetic tasks involving 32 Google Workspace APIs. The OOD tests retain the same tools and permission semantics, making it difficult to assess generalization to more complex real-world environments.

Capability exposure is measured by equally weighting accessible capabilities, without accounting for differences in their actual risks. Therefore, the reported 27% reduction in exposure does not necessarily imply a corresponding reduction in security risk.

Figure 1 is text-heavy and mainly lists procedural steps, rather than clearly illustrating the key mechanisms of value comparison and capability-aware selection.

Questions for authors

How does DARP compare with the main baselines in terms of task recovery and capability exposure when using Qwen-estimated values?

How does the method perform when recovery requires more than three steps or when candidate generation is incomplete?

Does DARP still outperform the baselines under a risk-weighted capability exposure metric?

2 comments
AuthorsOpenReview ↗

Response to Reviewer MKB4

Thank you for your review.

Q1. ... exact task values ... exposure ...

Answer: The new score-matched experiment holds the Qwen scorer, checker, execution budget, and initial decision inputs fixed. At the submitted ϵ=0.1\epsilon=0.1, DARP lowers exposure from 9.85 to 6.82 and raises FAA from 76.88% to 86.25%. On 280 recoverable tasks, both methods complete 249; only DARP completes 2, only Constraint Filter completes 16, and neither completes 13. DARP has lower exposure on 37 shared completions and equal exposure on 212. In the new ϵ=0.05\epsilon=0.05 run, exposure is 25.21% lower, FAA is 11.88 points higher, and SRR is 94.38% versus 95.31%. The revision adds the full grid, paired intervals, and task-level comparison in Figure 2, Table 15, and Table 16.

Q2. ... OOD ... tools ... permission semantics ...

Answer: The revised benchmark description separates the 160 unseen workflow compositions from the new AgentDojo evaluation. The former uses the same permission catalog and reaches 100% SRR with exact values under each of three candidate models. The latter runs the original Progent policy and AgentDojo executor on 12 independently authored tasks in four domains. On four matched, induced denials, both arms execute a legal next call in all four cases and finish two tasks. The revision reports these two settings separately in Section 4, Table 5, and Table 21.

Q3. ... weighting accessible capabilities ... risks ...

Answer: The revision defines exposure as newly reachable access coverage in Eq. (14) of Section 3. The new weighted analysis keeps the executed trajectories fixed and assigns read, write, and high-impact operations weights of (1,2,4)(1,2,4) or (1,3,9)(1,3,9). DARP's mean exposure is 31.59% and 33.92% lower than Constraint Filter's on the same 320 exact-value tasks, with 100% SRR for both. Table 18 gives the weights and results.

Q4. ... Figure 1 ... value comparison ...

Answer: The revised Figure 1 follows a recorded Gmail label update through the six stages of the submitted diagram. Two legal requests have value 0.925 and clear the 0.825 frontier cutoff. DARP selects gmail.modify as at∗a_t^* because its incremental exposure is 19 versus 22 for full mail access. The gateway verifies the selected request before dispatch, and the returned grant informs the next proposal. Table 14 gives the complete candidate comparison.

Q5. ... task recovery ... capability exposure ...

Answer: The revision adds the score-matched Constraint Filter comparison in Table 15. On the 320 ID tasks, Qwen-estimated DARP has 89.64% TRR, 90.94% SRR, and exposure 6.82. Constraint Filter has 94.64%, 95.31%, and 9.85, respectively. The same table reports Self-Revision at 96.43%, 95.94%, and 8.24, and Direct Retry at 84.29%, 85.63%, and 7.97, using their original revision rules. UDR is zero in all four runs.

Q6. ... three steps ... generation ... incomplete ...

Answer: The new constructed chain checks in Appendix I (Longer recovery chains, p. 20) complete the four-step chain at H=4H=4 and H=6H=6, and the six-step chain at H=6H=6. At shorter tested horizons, the chains select safe stop. Completed chains use the narrower request and certify each step. The revision also adds a 320-task ablation in Table 17. Model proposals plus safe stop reach 87.50% candidate recall and 98.13% SRR with zero UDR, using the same exact transition and value model.

Q7. ... risk-weighted capability exposure ...

Answer: Table 18 adds two impact-weighted profiles. DARP has lower weighted exposure than Constraint Filter at the same 100% SRR, and lower exposure than Self-Revision under both profiles. On each exact-value pair with Constraint Filter, DARP exposes each API method no more often and adds the same number of scope identifiers. This ordering holds for nonnegative method weights and a common scope weight. Table 19 also reports 27.95% and 30.21% lower weighted exposure on the 249 recoverable tasks that DARP and Constraint Filter both complete with Qwen estimates at the submitted setting.

Thank you for the response. The additional clarifications alleviate some of my concerns, but considering the paper’s quality and the response as a whole, I do not find sufficient grounds to raise my rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.