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.
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.