acceptodds
Under review as a conference paper at ICLR 2027

APEX: Active Protection at Execution Boundaries for LLM Agents

Abstract

Indirect prompt injection (IPI) hides adversarial instructions in content that large language model (LLM) agents read at runtime. As agents compose heterogeneous capability units, including Tools, MCP servers, and Skills, the carriers of injection multiply, and defenses built to recognize attack patterns fall behind them. We instead shift defense from covering attack patterns to one stable point: whatever the carrier and however the injection propagates, harm materializes only at the execution boundary, where the agent turns internal state into an external action or released output. Safety there turns on two conditions, both settled by the trusted task rather than by the run: whether the proposed effect is authorized, and whether the runtime information reaching it is endorsed by that task. We present APEX, an active defense that enforces both at this boundary from a single authorization contract compiled before untrusted execution: evidence-gated prevention admits an effect only when the contract justifies it, while deception-based exposure makes unendorsed use reveal itself before the effect commits. Protection therefore follows from what the task permits rather than from how an attack is built, and applies uniformly across capability units without attack-specific policies or taint tracking. Against 13 baselines, APEX attains 0% attack success on five of six benchmarks and 0.56% on the sixth, holds 0% under adaptive attacks on all three capability-unit types, and remains effective across defender backbones.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agent security, Active defense, Indirect prompt injection

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_BQ62OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces APEX, a defense against IPI in LLM agents that is meant to check actions proposed by an agent against an "authorization contract" compiled from the trusted user task. The authors focus on executions with an unauthorized effect or arguments (AauthA_{\mathrm{auth}}) and executions that use information through a dependency that the task does not endorse (AinflA_{\mathrm{infl}}). To stop executions in AauthA_{\mathrm{auth}}, they construct the authorization contract as a DAG, checking whether an agent-proposed action is permitted by this contract and recorded runtime evidence (WRAP/evidence-gated prevention). To stop attacks in AinflA_{\mathrm{infl}}, they use canary-style probes to detect when the agent uses information that the task did not authorize it to use (PLANT/deception-based exposure)

The authors evaluate APEX on 6 benchmarks against 13 baselines and report both low ASR and comparably maintained task utility. The evaluations also include ablations, adaptive attacks, and runtime overhead measurements.

Critical strengths
  1. The approach to IPI defense in agents seems useful and interesting. Focusing on the actions an agent proposes is a natural place to secure LLM agents, and checking them against the trusted user task makes sense. The authors propose a concrete system that combines authorization checking with canary-style probes and seems to protect well against different types of prompt injection.
  2. The experimental evaluation is strong. The authors evaluate APEX across 6 different benchmarks and against 13 baselines. The experiments give me confidence in the reported results for both ASR and task utility. I also appreciate the inclusion of adaptive attacks and ablations showing the contributions of WRAP and PLANT.
Critical weaknesses
  1. Clarity. The use of terms such as “capability units,” “boundary values,” and “endorsed dependencies” without a definition when they are introduced, nor any mention of how they might relate to existing concepts in agent security, makes it difficult to determine what is new and to follow the ideas being presented.

Important details are also left out of the main paper, which makes it hard to tell which parts of the defense are programmatic and which use LLMs. For example, Appendix B states that APEX coordinates three subagents, which is not clear from a reading of the methodology.

  1. Authorization contract reliability. The authorization soundness guarantee requires that the authorization contract accurately represent the user’s intent. This is a strong assumption that is not tested in the paper. Appendix I.1 argues that because the LLM that creates the contract does not read untrusted text, an incorrect contract might block legitimate work but will not admit harm. This seems wrong, since the contract could still be overly permissive, which would let WRAP allow an unauthorized action to be executed.

  2. Generalization to open-ended tasks. While this is partly addressed as a limitation and an assumption for APEX to work, it is still a weakness of the method. Agents work on multi-step, open-ended tasks, but because the contract is determined before execution, APEX will block any action that could not be anticipated from the user’s initial request.

  3. Insufficient evidence for WRAP’s “valid semantic bindings” assumption. The soundness proof relies on the assumption that when more than one piece of evidence could satisfy a Derive clause, an LLM picks the correct one. Similar to point 2, there is currently no evaluation of how often these judgments need to be made or of their accuracy. This seems particularly important since this LLM makes these judgments while reading untrusted content.

Questions for authors
  1. How do the authors know that the authorization contracts won't admit harm by being overly permissive? The 34 contract errors in Appendix I.1 seem focused on failures associated with an overly restrictive contract. Are there cases where the contract is ever overly permissive?

  2. How often does WRAP rely on LLM semantic bindings (such as selecting evidence for a Derive clause), and how accurate is it? Ideally, an adaptive attack targeting these bindings would be evaluated to help show whether this is a practical risk.

5 comments

We sincerely thank the reviewer for your valuable feedback and the effort you have devoted to reviewing our work. We fully understand that reviewing manuscripts requires substantial time and effort. Following your feedback, we have carefully revised the manuscript PDF and conducted a series of additional experiments to further strengthen our work and address your concerns. We would be sincerely grateful for your further review and guidance on our revisions.

We sincerely thank the reviewers for their valuable feedback and time. We have carefully addressed your concerns, revised the manuscript, and added supporting experiments. We would greatly appreciate it if you could revisit our responses and let us know if any concerns remain or further clarification is needed.

Dear Reviewer, We have made every effort to address your concerns. As we are approaching the end, we would appreciate your feedback on whether our responses have adequately addressed your concerns. If any issues remain, we would be happy to provide further clarification.

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

The paper proposes APEX, a runtime defense against indirect prompt injection (IPI) for agents that use heterogeneous capability units (Tools, MCP servers, Skills). The core idea is to stop enumerating attack patterns and instead enforce at the execution boundary, where an injection must become an external effect or released output. Before any untrusted content is read, an LLM compiles the trusted user task and a task-independent capability registry into an authorization contract: a DAG of Acquire / Derive / Conditional / Effect clauses. At runtime, two mechanisms read this contract:

  • WRAP (evidence-gated prevention) admits a proposed effect only if a matching Effect clause exists and every argument can be reconstructed from receipts and upstream clause outputs. Scripts enforce candidate membership and exact reconstruction; an LLM resolves ambiguous bindings.
  • PLANT (deception-based exposure) replaces identity-bearing spans of untrusted content (or adds synthetic elements or inert artifacts) with per-episode tokens. A proposal that carries a token witnesses unendorsed use.

Rejected proposals go through a non-agentic Safe Continuation step (repair / replan with a fresh agent / abort) under the same frozen contract. The paper gives conditional authorization-soundness and exposure-coverage arguments (Appendix B) and evaluates on six benchmarks: AgentDojo, ASB-OPI, MCPTox, MSB, SkillInject, and SCR-Bench. Against 13 baselines, it reports 0% ASR on five benchmarks and 0.56% on SkillInject with high utility, 0% ASR under three adaptive attacks, and similar security with three defender backbones.

Critical strengths
  1. Clear and principled enforcement criterion. Splitting boundary risk into unauthorized effects (A_auth), authorized effects using unendorsed information (A_infl), and an explicitly out-of-scope residue (U_q) is clean. It also makes clear what the defense does not claim (e.g., the SkillInject content-bias case in Appendix I.2).
  2. Strong and broad empirical results. Across six benchmarks spanning three capability types and composition attacks, APEX has the lowest ASR, often with higher attack utility than the undefended agent (e.g., ASB-OPI AU 71.23→85.05 at 0% ASR; MCPTox AU 41.25→75.52). On SCR-Bench composition attacks, where every baseline either fails (TaskShield 68.97%, Progent and DynamicGuardian ≈89% ASR) or collapses utility (ClawGuard), APEX reaches 0% ASR with 99.85% AU.
  3. Determinism where it matters. Conditional operators, exact reconstruction, and the continuation controller are script-enforced, and the LLM can only choose among script-admitted candidates. Continuation never modifies the contract, and every repaired proposal is re-checked.
  4. Transparent reporting of mechanism reach and failures. Table 5 reports PLANT deployment, abstention (83%), and trigger rates (7.6%); Table 9 attributes all 813 attack-case utility errors. Baseline configurations are pinned (e.g., StackOne commit, MELON threshold, CaMeL ASB policy frozen against a catalog hash). Code is provided via an anonymous repository.
Critical weaknesses

W1. The safety-direction reliability of the LLM-dependent components is unmeasured, so the guarantees are only as strong as untested assumptions. Authorization soundness (Appendix B.2.2) rests on three assumptions: contract fidelity, correct receipts and scripts, and correct semantic bindings when matching is ambiguous. The task-level theorem additionally requires that every external effect is mediated. Each of these is supplied by an LLM or a heuristic:

  • Contract fidelity. The contract is compiled by the defender LLM. Appendix I.1 argues that contract errors "block legitimate work rather than admitting harm," but Table 9 only counts utility failures. An over-permissive contract (e.g., an Effect clause with a loosely constrained Derive relation, or an extra recipient position) would be admitted faithfully by WRAP and appear neither as a utility error nor necessarily as ASR on these benchmarks. A manual audit of a stratified sample of compiled contracts for over-authorization, per benchmark and per defender backbone, is needed to support the one-sided-error claim.
  • Semantic bindings. For Derive clauses with open relations (e.g., c1 in Figure 7: "Python code implementing the requested registration and login functionality"), essentially any content satisfies δ_c. In the J.2 case, WRAP admits the backdoored script, and protection comes entirely from PLANT having tokenized the two literal credentials. The paper's claim that "substituting other values calls for a derivation in C_q that no clause supplies" seems to contradict the preceding statement that c1 constrains content only by its declared relation. If c1 admits any login implementation, it should also admit one with a different hard-coded admin account. Please clarify what δ_c1 and its candidate constraints are.
  • Effect classification. For MCP, χ_eff is assigned by a name rule ("externally consequential unless the name marks it as a read-only retrieval"), and only χ_eff = 1 units are mediated. Many "read-only" retrievals carry attacker-chosen arguments and can exfiltrate (e.g., fetching a URL with data in the query string, or search/lookup calls to attacker-controlled endpoints). Please report how many units in each catalog are classified as non-effect, whether any accept free-form arguments that reach external parties, and the effect of setting χ_eff ≡ 1 for all MCP units (which, by the paper's own argument, should only cost extra checks).

W2. The adaptive evaluation does not attack APEX's mechanisms, and the zero-ASR claims need bounded interpretation. The three adaptive attacks (AutoDojo, MCP-ITP, SkillJect) are generic feedback-driven optimizers with small budgets: at most 4 candidates for AutoDojo, 2×5 for MCP-ITP, and 5 attempts for SkillJect. They are run on 400 / 455 / 44 applicable cases. For AutoDojo, candidates are ranked first by benign completion and only then by attack success, which can discard attack-successful candidates (Table 2 shows undefended AU rising from 83.25 to 97.00 under "adaptation"). More importantly, none of these attacks is designed with knowledge of APEX. Obvious APEX-aware strategies include:

  • making the injected identity not appear as a single copyable span (e.g., "the username is dev followed by _admin", spelling out characters, or encoding it), so that PLANT's span replacement does not carry the token while the agent reconstructs the value;
  • steering the agent to obtain the attacker value through an endorsed channel (e.g., inducing a lookup whose result legitimately binds to an Acquire clause);
  • exploiting loose Derive relations or non-mediated "read-only" units (W1);
  • attacking the contract compiler indirectly through the capability descriptions it reads during registration/compilation.

Without such an evaluation, "holds 0% under adaptive attacks" (abstract) overstates the evidence. In addition, AgentDojo has no adaptive result: Appendix F says AutoDojo provided limited improvement there and omits it.

W3. Several ablation and backbone numbers are inconsistent with the stated case counts, and no result reports uncertainty.

  • Table 3 (ablation) has values that cannot arise from the stated counts. On SkillInject (180 attack cases), w/o-PLANT ASR 6.45 (11/180 = 6.11, 12/180 = 6.67), w/o-WRAP ASR 2.23, and AU 60.22 / 68.72 are not of the form k/180, and BU 75.72 / 71.51 / 75.76 are not attainable on 90 or 180 benign cases. On ASB-OPI, w/o-Continuation AU 62.20 is not k/2040.
  • Table 6 has similar inconsistencies: AgentDojo 71.51 and 82.88 (n = 629), ASB-OPI 85.55 (2,040), MCPTox 80.92 (1,348), MSB AU 97.03 / 97.12 (attainable on neither 415 nor 622), MSB ASR 0.56 (622), and SkillInject 88.60 (180).
  • Possible explanations are subsets, failed runs, or averaging over repetitions, none of which is described. Since Table 3 is the only causal evidence for the two-primitive design and Table 6 supports "effective across defender backbones," please report k/n for every cell and state the per-arm population.
  • The paper reports no confidence intervals, repetitions, or decoding settings ("temperature" does not appear). With LLM-based compilation, binding, placement, and GPT-5.4 judging, run-to-run variance may be non-trivial. Several "lowest ASR" claims rest on single-case differences: AgentDojo 0/629 vs. Spotlighting 1/629; MCPTox 0/1,348 vs. MCPGuard 1/1,348; SkillInject* 0/44 vs. TaskShield 1/44. Zero counts also have meaningful upper bounds: by the rule of three, ~0.5% on AgentDojo and ~6.6% on the 44-case SkillInject* subset. Please report intervals and, at least for a subset, repeated runs. The GPT-5.4 judge, which decides ASR from free-text artifacts on SkillInject/SCR, should be validated against human labels.

W4. The contribution of PLANT, a headline mechanism, is modest and benchmark-dependent, which the abstract and Figure 1 do not convey. On ASB-OPI, the largest benchmark (2,040 of 5,486 attack cases), removing PLANT leaves ASR at 0.00, so all protection there comes from WRAP. Table 5 reports 779 triggered probes across all benchmarks. Since triggers are counted per probe instance, at most ~14% of attack cases involve any PLANT witness. Probe deployment covers only 41% (MSB) and 38% (SCR) of cases. The ablation omits SCR-Bench, the benchmark with the highest trigger rate (64%). PLANT clearly matters on MCPTox and SkillInject (ASR 0 → 6.45 without it), so this is not a fatal issue. But the framing of two co-equal primitives should be qualified, and a per-case attribution (rejected by WRAP predicate vs. PLANT witness vs. both) should be added for all six benchmarks.

W5. Clarity and consistency issues.

  • Placement rule: §3.3 says probes "go on the endorsed dependencies, never on a value the task requires," whereas the proof in B.3.2 says "probes are placed outside the endorsed dependencies of the protected effect." Please state one rule and explain which property (zero false positives vs. coverage) depends on which reading.
  • "Two strongest baselines" (§4.3) is undefined. For ASB-OPI, Table 2 uses MELON and CaMeL, although DRIFT has the lowest ASB-OPI ASR in Table 1 (9.80). "Dynamic Gradient" in Table 2 should presumably be Dynamic Guardian.
  • §4.3 promises results for the remaining cases in Appendix G, but G contains only construction and selection rules. Appendix F calls MSB* AU of 61.32% (down from 82.55%) "maintaining high utility."
  • Fig. 4's SCR aggregates for ClawGuard/Progent use 551 cases while the others use 667 (Table 8 footnote); this should be marked in the figure.
  • PLANT rewrites untrusted content the agent sees (e.g., IBANs or credentials replaced by CERT-… tokens). The paper should discuss when this alters benign task semantics beyond the "no PLANT false positive" count, e.g., when the user legitimately needs the replaced value in the final response.

W6. Missing comparisons with the closest prior work and incomplete cost reporting.

  • The related work positions APEX against contract/graph/provenance defenses that also derive a specification from the user task: IPIGuard, FAVA, AuthGraph, AgentArmor, CXI, PACT, and AttriGuard. None of them is a baseline, while most compared methods are detection or policy defenses. At least IPIGuard (public) or one graph/provenance method should be included on AgentDojo/ASB-OPI. Otherwise the main claim, that binding argument derivation in a fixed contract is what yields 0% ASR, is not separated from what a planned-dependency-graph defense already achieves.
  • On AgentDojo with DeepSeek-V4-Flash, the undefended ASR is only 17%, and Spotlighting attains 94.85 BU / 86.96 AU / 0.16% ASR, i.e., higher utility than APEX at essentially the same security. This benchmark therefore provides limited discrimination; a more injection-susceptible task agent would be more informative.
  • Figure 5 reports cost only on three benchmarks and only graphically. APEX appears to be among the most expensive methods on MCPTox and SkillInject (the SkillInject axis extends to ~10M tokens). Please give exact per-task tokens, LLM calls, and wall-clock latency relative to the undefended agent for all six benchmarks, broken down into compilation, WRAP binding, PLANT placement, and continuation.
  • Development/test separation is not described. The operator set (Table 4, e.g., interval free, aligned lookup), the registry rules, and the prompts were presumably developed with these benchmarks in view. Was any benchmark held out during development?

Reproducibility note. I appreciate the anonymous repository. I downloaded the full archive (131 files). It contains the defender, the contract-compilation skill prompt, registry manifests, packaged case descriptors for the six benchmarks, integrations for all 13 baselines, and unit tests. It contains no experiment logs, result files, or per-case compiled contracts or traces, consistent with its README/ANONYMITY.md, even though the reproducibility statement says "contracts and evaluation artifacts are available." Releasing the per-case contracts, decision traces, and judge outputs, together with decoding settings, would make the W1/W3 questions checkable.

Questions for authors
  1. Can you audit a stratified sample of compiled contracts (per benchmark and per defender backbone) for over-authorization, i.e., effects or argument sources the user did not authorize, and report how often admitted proposals relied on semantic (LLM) bindings rather than exact matching?
  2. In the J.2 case, what are δ_c1 and its candidate constraints? Would WRAP admit a script containing a different hard-coded admin account not present verbatim in the skill?
  3. How many MCP units are classified χ_eff = 0 by the name rule? Can any of them send attacker-chosen arguments to external parties? What happens to ASR/AU/cost with χ_eff ≡ 1?
  4. Can you evaluate an APEX-aware adaptive attacker that knows the contract/probe mechanism (e.g., splitting or encoding identities, routing attacker values through endorsed Acquire channels, exploiting loose Derive relations), with larger budgets and attack-success-first ranking, including on AgentDojo?
  5. Please give k/n for every cell of Tables 3 and 6 and explain how the non-attainable values arise. Please also report confidence intervals, decoding settings, and at least a few repeated runs, plus a human-agreement check of the GPT-5.4 judge.
  6. What is the per-case attribution of blocked attacks to WRAP vs. PLANT on all six benchmarks, and the PLANT ablation on SCR-Bench?
  7. How does APEX compare with IPIGuard (or another graph/provenance defense) under the same backbone?
  8. How is the PLANT nonce generated in the experiments? The code requires a fixed, non-random per-episode value, while the paper states tokens are unpredictable to the adversary.
3 comments

We sincerely thank the reviewer for the valuable and constructive feedback on our work. We understand that reviewers are handling a substantial workload and greatly appreciate your time and effort. Following your suggestions, we have carefully revised the manuscript and conducted additional experiments to further support and clarify our findings. We would be sincerely grateful for your further review and guidance on our revisions.

We sincerely thank the reviewers for their valuable feedback and time. We have carefully addressed your concerns, revised the manuscript, and added supporting experiments. We would greatly appreciate it if you could revisit our responses and let us know if any concerns remain or further clarification is needed.

Dear Reviewer, We have made every effort to address your concerns. As we are approaching the end, we would appreciate your feedback on whether our responses have adequately addressed your concerns. If any issues remain, we would be happy to provide further clarification.

Reviewer_631MOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes APEX, a defense against indirect prompt injection that checks agent actions before they produce external effects. At a high level, APEX first compiles the trusted user request into a fixed authorization graph (DAg) specifying permitted actions and the evidence that may determine their arguments. Then it checks proposed actions against this graph through evidence-backed reconstruction, while inserting probes into selected runtime content to expose unendorsed information use. Rejected actions trigger repair, replanning, or termination without changing the original authorization. Across six benchmarks covering Tools, MCP, and Skills, the authors report zero attack success on five benchmarks and 0.56% on the remaining one.

Critical strengths
  • The strongest aspect is the use of a shared authorization contract to connect argument-level checking, probe placement, and recovery. This gives the components a common basis for deciding which information may influence an action and what evidence remains usable after a rejection.
  • Paper provides sufficient ablations to defend the design choices with evidence that the proposed components address different failure modes. In particular, removing WRAP raises ASR on ASB-OPI from 0% to 57.60%, whereas removing PLANT raises ASR on MCPTox from 0% to 6.45%
Critical weaknesses

I think the major gap is that the reliability of the semantic decisions underlying the authorization guarantee is insufficiently evaluated. In the implementation, an LLM decides which evidence satisfies a Derive relation, scripts check source membership, permitted composition, and exact reconstruction. While these checks constrain the model’s choices, an exact value from an allowed source can still be the wrong value for the intended purpose. The paper therefore needs a direct evaluation of incorrect bindings and overly permissive contracts.

Questions for authors

I would encourage to connect the work more strongly with the existing literature especially in formal verification. For example, paper discusses FAVA and AutoGraph but there are absent from the experimental comparisons.

2 comments

We sincerely thank the reviewer for your positive assessment and recognition of our work. We have carefully revised the manuscript PDF and tried our best to address your suggestions in our response. We would be sincerely grateful for your further review and guidance on our revisions.

We sincerely thank the reviewers for their valuable feedback and time. We have carefully addressed your concerns, revised the manuscript, and added supporting experiments. We would greatly appreciate it if you could revisit our responses and let us know if any concerns remain or further clarification is needed.

Comments on the paper

1 comment

Dear Chairs/Reviewers,

Thank you for handling our submission. We summarize how our revised manuscript addresses the three reviewers' concerns. All changes are marked in blue.

Main improvements

Our revisions focus on three questions central to APEX's contribution:

1. How reliable are the authorization decisions? We directly audited contract compilation and semantic binding, which were important concerns raised by all three reviewers. Among 1,747 contracts, 16 (0.92%) were over-permissive and 8 (0.46%) over-restrictive. The over-permissive contracts covered 72 attack episodes, none of which succeeded. For semantic binding, 396 of 398 auditable accepted bindings (99.5%) preserved their intended task roles. We also clarified the assumptions and limitations of these LLM-dependent components in §3 and §5.

2. Do WRAP and PLANT provide independent protection? We added per-case attribution across all six benchmarks: WRAP exclusively blocked malicious proposals in 2,573 cases, PLANT exclusively blocked them in 330, and both triggered in 338. Additional SCR-Bench ablations show that removing PLANT increases ASR from 0% to 24.14% on AuthBlur and 2.67% on CapFlow, establishing its contribution beyond WRAP.

3. Does the protection remain effective under stronger comparisons and targeted attacks? We added official IPIGuard and AttriGuard baselines, mechanism-targeted boundary stress tests, and sensitivity analyses of capability registration. APEX retains 0% ASR across five of six benchmarks in the original evaluation, while the new tests examine attacks against its specific authorization and probe mechanisms.

How the review concerns were addressed

ConcernReviewersNew evidence or revisions
Contract fidelity and over-authorization631M, xKdt, BQ62Audit of 1,747 contracts, including over-permissive and over-restrictive cases and their downstream outcomes (§B.2.1, Tables 5, 18)
Reliability of LLM semantic binding631M, xKdt, BQ62Audit of 436 accepted bindings, with 396/398 correct auditable cases; targeted tests of authorized evidence and open Derive relations (§B.3.2, §H.2, Tables 8, 16)
Methodology clarity and implementation responsibilitiesBQ62, xKdtRevised §3 with a worked example based on Fig. 2, clearer WRAP/PLANT execution steps, and explicit distinctions between LLM reasoning and script-based checks
APEX-aware attacks and authorization boundariesxKdtTargeted stress tests of identity obfuscation, probe evasion, evidence laundering, and open Derive relations; attacker-tool registration sensitivity (§H.2, §B.2.2, Tables 6, 16)
Independent contribution of PLANTxKdtPer-case attribution and SCR-Bench PLANT ablation (§C, §B.4.1, Tables 10, 11)
Closely related baselines631M, xKdtOfficial IPIGuard and AttriGuard comparisons under the same backbone (§4.2, §G, Table 1)
Statistical consistency and reproducibilityxKdtCorrected ablation and backbone tables with exact counts and confidence intervals; three independent runs on 10% of each benchmark (§4.4, §D, §M)
Effect classification, efficiency, and scopexKdt, BQ62All-effect MCP sensitivity, six-benchmark cost breakdown, and clarified limitations of fixed contracts and tasks requiring reauthorization (§B.1.2, §E, §5)

Scope of the guarantees

We have also strengthened the discussion of limitations. Contract compilation may incorrectly expand or restrict authorization, while semantic binding remains dependent on the correctness of LLM judgments over potentially untrusted evidence. Our revised §5 makes these limitations explicit, and Appendices B.2–B.3 evaluate their observed reliability. We further clarify that tasks requiring new permissions beyond the initial trusted request need trusted reauthorization.

We respectfully hope the revised manuscript and reviewer discussion will inform your final assessment. Thank you for your consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.