- Rating
- 1
- Confidence
- 3(−1)
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
The paper studies authorization for tool-using LLM agents during failure recovery. The paper shows that per-call guards can evaluate each recovery action in isolation and therefore miss cases where earlier effects have already consumed part of the original grant. The paper proposes Recovery-Closed Authorization (RCA), keeps a trusted ledger of realized effects and allows recovery actions only if their effects stay within the remainder.
Critical strengths
The problem is real and well motivated. Stateless per-call guards may indeed fail to catch failed tool calls or compositional effects.
Pushing such decisions to a separate security layer is practical and well founded in prior work.
Critical weaknesses
Without the excess jargon (of which there is a lot, for a well known idea), the proposed RCA is a type of cumulative quota enforcement. Tracking a running total of consumed authority and rejecting actions that would exceed the limit is a relatively known idea standard in payment systems and rate limiting and the paper should position itself against that literature rather than only against recent LLM agent systems. From the perspective of that body of work, there are several hard parts of deploying such quotas, such as tracking realized effects reliably or handling authorization constraints that aren’t additive. However, in this work, all of these are placed inside the trust assumptions. The paper risks translating standard methods under generous assumptions to agentic systems without any specific insights about agentic systems themselves. Given this is not a particularly new idea, I would also heavily advise simplifying the coined jargon, e.g. "declared effect abstraction", "observed prefix authorization violations", "heterogeneous current-step mechanisms", which makes the abstract and paper difficult to understand.
Critically, the numbers in table 1 and 6 do not match (S2 is 52.8% in Table 1 but 14.33% in Table 6 for model candidate). I have significant concerns about the soundness of headline results.
The evaluation is also circular. The oracle is defined along the same frame as RCA, so complete agreement is somewhat tautological. The baselines also seem to lack the exact states RCA uses or are incomplete comparisons. Why does LedgerLocal ignore cumulative totals?
The theoretical results are somewhat overstated. Theorem 1 states that a deterministic monitor receiving identical inputs in two histories cannot make different decisions. This is again tautological. Deterministic functions must have a single output for each input by definition.
RCA usually stands for root cause analysis and I would advise changing the acronym. This does not affect my score.
Questions for authors
How does a simple cumulative-budget rule written in AgentSpec or Progent compare to RCA?
What are the actual measured quantities?
5 comments
I thank the authors for the detailed response and experiments. The SQLite execution study and the fixed-candidate replay are useful additions.
However, the revised manuscript and rebuttal actually reinforces my concerns about the contribution. Line 1428 now reads "A correct per-root amount counter is sufficient for the amount-only accounting problem" and that “RCA does not improve that decision”. The next sentence says the remaining episodes "are attributable to cumulative count", line 1429, which the manuscript notes “is itself additive”. The manuscript seems to now imply that RCA is functionally equivalent to cumulative quota enforcement over both amount and count. By the revised paper's own argument and even the rebuttal ("Accordingly, we do not claim that RCA is intrinsically more expressive than a complete multidimensional history-aware policy. Indeed, documented-hook implementations supplied with equivalent trusted state and locking can reproduce RCA's decisions.") it seems they would match in performance. I do not see positioning against such quota systems in spite of this concession.
On the topic of Table 1 and Table 6, I am glad that the inconsistency was resolved. However, tables 21 and 22 on Progent and AgentSpec are identical in all reported statistics. Is there a theoretical reason these are meant to be identical? If the choice of framework does not influence the actual results, I struggle to see in what sense these frameworks are considered evaluated.
My understanding is now that the paper is proposing a method that it then concedes is functionally equivalent to a cumulative quota enforcement. Of the two theorems to demonstrate the contribution, one is now conceded as straightforward in rebuttal. The oracle evaluation is now “an upper-bound comparison within the same declared abstraction, rather than an independent validation of the abstraction itself". I will keep my score.
Clarifying RCA’s Problem Novelty, Technical Contributions, and Cross-Framework Evidence
We thank the reviewer for the follow-up. However, we respectfully disagree with the conclusion that RCA's functional equivalence to sufficiently expressive stateful enforcement implies a lack of research novelty. This reasoning conflates the expressibility of a security property with the originality of identifying, formalizing, and systematically evaluating that property.
1. Functional equivalence does not invalidate the problem contribution
RCA's central contribution is the identification and formalization of failure-induced authority drift as a distinct recovery-time authorization problem.
The critical observation is that two histories can expose the same failure observation and the same recovery candidate, yet require opposite authorization decisions solely because their previously realized effects differ.
This problem is not restricted to replayed requests, expired permissions, or duplicated authorization tokens. A newly generated recovery action may have valid credentials, satisfy the current-call policy, and use a fresh transaction identifier, yet exceed the authority remaining after earlier partial or successful execution.
The existence of a sufficiently expressive policy language capable of implementing the required check does not establish that the underlying security problem, its decision-state requirements, or its recovery-closure invariant were previously identified and characterized.
Indeed, the equivalence demonstrated by our documented-hook experiments establishes that existing authorization frameworks can implement the RCA contract when given the necessary trusted information and synchronization. This supports the portability of the proposed security property; it does not negate its formulation.
Our technical contributions should therefore be evaluated at the level of the security problem and its formalization:
- Problem: Failure-induced authority drift, where locally indistinguishable recovery requests require history-dependent authorization decisions.
- State characterization: Residual future-effect decisions identify the information that sound and permissive recovery authorization must preserve, relative to the declared effect abstraction and continuation family.
- Security invariant: Recovery closure constrains every prospective effect against authority remaining after realized execution, including under uncertainty.
- Composition: The invariant can be enforced alongside existing invocation-time authorization without changing their native decision rules.
Theorem 1 establishes the necessary distinction between authorization histories; Theorem 2 establishes the resulting recovery-closure guarantee under explicit trust assumptions. The originality claimed is the recovery-specific security formulation and its enforcement contract, not the complexity of the induction technique.
2. RCA studies a different authorization question from the cited work
We agree that RCA shares durable-state and authorization-checking mechanisms with CapLease [1] and commit-time authorization [2]. However, sharing implementation ingredients does not make their security objectives identical.
CapLease [1] studies the repeated consumption of token-independent authorization through reissuance and semantic replay. Commit-time authorization [2] studies whether an authorization justification remains valid and bound to the effect at the point of durable commitment.
RCA asks whether a proposed recovery remains authorized after the composition of previously realized effects has changed the authority available for further execution.
This distinction matters when the next action is neither an exact replay nor dependent on stale authorization evidence. The recovery action itself can be locally valid while its composition with prior effects violates the grant.
Our contribution is to isolate this failure mode, characterize its decision-relevant state, and evaluate the security–utility consequences of enforcing recovery closure. We do not claim that the neighboring systems could never be extended to enforce the same property.
Beyond Per-Call Authorization—Fair Quota Baselines and Verified Recovery Effects
4. Theory, trust assumptions, and implementation costs
We agree that the elementary indistinguishability argument underlying Theorem 1 is straightforward. Its purpose is not to claim a new general theorem about deterministic functions, but to identify which execution-history distinctions an authorization monitor must retain to make sound and permissive recovery decisions. The accompanying characterization is scoped to a declared effect abstraction and separating continuation family.
Similarly, Theorem 2 establishes a conditional recovery-closure invariant. It does not establish that arbitrary external APIs automatically provide faithful envelopes or trustworthy effect observations.
The implementation contract makes those requirements explicit: trusted effect accounting, conservative prospective envelopes, mediated execution, and per-root linearization. RCA addresses authorization within that trusted boundary, rather than claiming to solve arbitrary external-effect instrumentation.
We also report the measured runtime costs (Appendix I). In a single-worker CPU microbenchmark, p95 checks were approximately 24.9 μs for native per-call authorization, 41.9 μs for scalar accounting, and 1,868 μs for the RCA ledger path. These measurements describe the evaluated implementation, not production-scale concurrency or end-to-end latency.
We appreciate the recommendation to simplify terminology. The revised manuscript distinguishes ordinary per-call authorization, cumulative accounting, and recovery-closure enforcement more explicitly, and identifies the measured populations and metrics directly in the corresponding tables. We also acknowledge the acronym overlap with root cause analysis.