acceptodds
Withdrawn from ICLR 2027

ResidualAuth: What Authorization State Must Agent Systems Preserve under Revocable Delegation?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

With revocable delegation, two histories can yield identical current permissions yet require opposite decisions for the same query after the same revocation. We introduce ResidualAuth, a theory-grounded framework characterizing the authorization state agent systems must preserve and evaluating its maintenance and use. Its formal core, residual authorization state, equates histories exactly when every future sequence of grants, revocations, and uses is valid after both or neither. We prove that exponentially many distinct residual states can nevertheless agree on who can reach whom through delegation paths. The analysis also yields exact or tight memory bounds as delegation redundancy varies and an average decision-error lower bound under an explicit bound on retained information. ResidualAuth evaluates information access, online state maintenance, and information use through a restricted executable benchmark and separate diagnostics. The benchmark pairs episodes differing in authorization-relevant history and requiring opposite decisions; pair accuracy requires both answers to be correct. To test use of supplied decisions, four open-weight models received trusted current-query Allow/Deny decisions alongside deterministic 256-token event extracts, achieving 15–16/16 correct pairs versus 0–2/16 with extracts alone. Memory diagnostics identified invalid reconstructions; models also answered incorrectly from valid memories that passed fixed future authorization tests. Together, these results distinguish what future authorization requires a system to retain from whether agents can access relevant information, maintain state across updates, and use available information to make correct decisions.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Language Agents, Authorization and Access Control, Revocable Delegation, Authorization Provenance, Agent Security

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_hmJZOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper investigates what authorization state an agent system must retain under revocable delegation. Its core observation is that two delegation histories can produce identical current permissions and reachability while requiring opposite decisions after the same revocation. The authors formalize this with Myhill–Nerode residual classes, give a rate–distortion lower bound on error under an information constraint, and introduce a paired benchmark ResidualAuth.

Critical strengths

The motivating example in Figure 1 is clear and the underlying point is meaningful. The pair-complete accuracy metric is a good design choice and the statistical testing rigor is good.

Critical weaknesses

The main text is extremely difficult to follow. Nearly every result is qualified with disclaimers about what it does not show. Several distinct protocols are introduced with little setup, relying on a "contract" in Appendix F which breaks self-containment of the main text, and important definitions (the selected-lineage contract, the deletion-sensitive design) are used as defined terms throughout but only summarized weakly in main text. From the main text alone it is hard to identify the primary empirical claim or how the studies fit together.

My below weaknesses are based off of my best interpretation of the main text.

The headline empirical results seem tautological. When a security layer supplies the correct Allow/Deny decision in context the model yields near-perfect accuracy. This reduces to either a study of whether the model trusts the security layer (potentially interesting and unexplored in main text) or whether the model can properly manage its context to recite the in-context answer (seemingly trivial). The hard gate is defined as E = A·Y, so its blocking every unauthorized effect follows by definition. Much of Figure 2 and the abstract is devoted to these outcomes.

The theory and experiments are disconnected. The paper states that token caps do not correspond to the information budget in Theorem 3, so the experiments do not test any of the bounds. The theoretical contribution itself relies on fairly standard tools and the main text does not make clear what is technically new beyond applying these to delegation graphs.

The empirical scale is small, with 16 pairs per cell for the main interface results and single seeded runs. When considering that the pairs come from fixed constructions the effective sample size is probably even smaller.

In general the figures are overly AI-styled. Figure 2 panel c is a whole quadrant of the figure that displays exactly two numbers. As noted above part d derives from the definition. Spacing of the legend and the titles would benefit readability. It's also not clear why certain figures only show certain models but not others, e.g. Figure 2 vs Figure 3.

Questions for authors

What do the authors consider the main technical contribution and how does it go beyond prior work on delegation and revocation?

In what ways is the empirical result counterintuitive or surprising beyond enacting the policy of a security layer?

Reviewer_L3mXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies what state an agent system must preserve to handle authorization and revocation queries. Its main observation is that two delegation histories can give the same end state, while requiring different authorization decisions after the same revocation. The authors derive bounds on the memory required under different delegation regimes and introduce ResidualAuth, a benchmark of authorization scenarios designed to test whether LLM agents follow delegation instructions. The results suggest that agents struggle to keep track of authorization reliably, and that hard enforcements can mitigate these failures.

Critical strengths

None

Critical weaknesses
  1. Much of the mathematics is trivial, and only present for the sake of itself. Theorem 3, in particular, basically states that if a system has insufficient information to reconstruct a randomly selected bit, there is a lower bound on prediction error. This is true, but it's unrelated to whether LLMs can update on authorization records. More broadly, the empirical analysis is not connected to the theory. For example, the paper derives memory bounds in terms of the number of principals, rights, and permitted grantors. However, the experiments do not test these bounds or vary the parameter that governs how much redundant delegation occurs.
  2. The large gains from authenticated authorization are unsurprising because the authors supply the correct decision for the query. Similarly, the fact that an authorization check prevents unauthorized effects is basically true by construction.
  3. The writing is confusing, often obtuse, and very Claude-ish. At no point do the authors explain the significance of the results or provide background
Questions for authors
  1. What is the novel insight, from a mathematical point of view, beyond applying combinatorial counting to delegation graphs?
  2. How does ResidualAuth differ from EAL-Bench in terms of empirical methodology and conclusions? The paper seems to understate their overlap
  3. What are the empirical consequences of the redundancy-memory law?
  4. Can the authors provide the source code to reproduce the experiments?
Reviewer_ModwOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies AI agents that pass permissions to one another, and shows that knowing who has access now isn’t enough. Revoking permission later depends on who gave permissions to whom. The study mathematically formalises how much of the permission granting must be handled and remembered by the system to always decide correctly and provides a benchmark framework to test various scenarios that look the same today but require opposite decisions after permission is removed. Results show that LLMs did well when handed correct permission records but worse when keeping track of it themselves.

Critical strengths

The authors focus on an interesting problem

Critical weaknesses
  1. The text lacks clarity. The domain is very abstract and reader needs to try hard in order to understand what is the problem statement.

  2. The paper is very long, many claims depend on reading an extensive appendix, and the experiments are very narrow in scope.

  3. Experimental design and results aren't clear. The study uses commercial LLMs on synthetic data with little to no explanation of its contents.

  4. Results show that LLMs perform better when they have access to graphs instead of keeping track of them by themselves, which doesn't offer very valuable insights by itself.

  5. Little novelty and Limited impact. Since the results are based on specific LLMs, the results have little to no value for the future. They could very easily become superseded one to two years from now when newer versions of Gemma, Qwen, and other LLMs are publicly available.

Questions for authors

What value do you believe this study offers to the wider research community?

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

Related papers

Discussion (0)

Sign in to comment.