acceptodds
Under review as a conference paper at ICLR 2027

From Safety Memory to Enforced Policy: Continuity and Scope Across Agent Replicas

Abstract

Adaptive agent defenses learn safety rules from past interactions to prevent recurring harm. Yet remembering a rule does not ensure that later actions obey it. We study this gap in serverless agents, where learning and execution can occur on different replicas. We propose Service-Global Monotonic Immunity (SGMI), which shares accepted rules across replicas, defines their action scope, and blocks matching tool calls before generative reconsideration. In Knative scale-out experiments, shared memory retrieves the learned rule on every fresh-replica probe but still permits 47.9% of harmful actions. SGMI blocks all 144 tested harmful probes across original and fresh replicas. In a paired banking ablation, recipient boundaries reduce legitimate-action rejection from 62.5% to 25%, while preserving all eight harmful-action blocks. SGMI connects adaptive safety learning with persistent, action-level protection across changing execution instances.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, safety memory, serverless computing

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_dL5BOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how safety rules learned by a tool using agent can remain available and enforceable after serverless replica replacement. SGMI stores accepted blocking rules in ordered shared state, maps proposed actions to structured predicates, and checks matches in a trusted tool wrapper. The distinction among visibility, scope, and application is useful. However, the current implementation and evaluation do not establish the claimed action level protection or a sufficiently general research contribution.

Critical strengths
  1. The visibility, scope, and application decomposition is clear and helps diagnose why merely sharing memory is insufficient.

  2. The paper recognizes that enforcement must bind concrete action arguments and versions rather than rely on another model judgment.

  3. The same scope policy control is a good experimental idea. The appendix is unusually candid about missing records, exclusions, implementation revisions, and inconclusive results.

Critical weaknesses
  1. The technical novelty is limited. A linearizable shared store, deterministic predicate matching, complete mediation, and single use authorization are established security mechanisms. The formal proposition follows directly from assuming correct matching, immutable prefixes, bound permits, and complete mediation. The paper does not convincingly show what SGMI adds beyond applying a conventional reference monitor to learned rules. The same scope policy obtains the same main safety and utility results, which further weakens the claimed distinction.

  2. The evaluation does not demonstrate the central claim in realistic agent behavior. The headline 144 blocked probes are repeated observations within twelve lifecycle cycles, not 144 independent attack settings. The main controlled banking study uses one archived rule and only eight harmful and eight legitimate calls. More importantly, every method, including no defense, has zero attack success in the ninety autonomous episodes. None of the 44 calls inspected by SGMI directly matches a rule, and SGMI blocks no action. Thus the autonomous experiment provides no evidence that SGMI protects an agent when an attack actually induces a harmful proposal.

  3. The evaluated implementation does not satisfy the contract used for the main claim. Appendix C reports that the retained wrapper executes a call after its arguments are changed and does not bind a changed read version. Only a subsequent revision rejects these cases. The lifecycle results retain the earlier request level interpretation, while the fixed sequence study queries a decision endpoint and executes admitted calls in a separate caller runtime. Consequently, the reported experiments do not validate the action binding and complete mediation assumed by Equation 3.

  4. Scope construction is manual, narrow, and brittle. The email projection uses domain specific text patterns and labels every modeled external send as occurring before review, even when the request states otherwise. The banking projection uses initial recipient history, which the paper correctly notes is not authorization. There is no meaningful evaluation of predicate induction, adversarial aliasing, alternative tools, conflicting or incorrect accepted rules, revocation, or epoch transitions. The guarantee assumes both rule correctness and that every covered action matches, leaving the hardest safety problem outside the result.

  5. Several evidence quality issues reduce confidence. The model checkpoint is unpinned. Full bodies for the 36 lifecycle records are unavailable. Some replica identities were read from the wrong field, initial utility states were not saved, and several cohorts or tasks were excluded because of implementation problems. The latency study uses one small local testbed and reports missing runs and 17 timeouts among 108 miss requests at parallelism twelve.

Questions for authors
  1. Can all safety experiments be rerun with the revised wrapper that binds frozen arguments, attempt identity, and read version?

  2. What measurable advantage remains over a conventional same scope reference monitor when both receive the same rule and predicate?

  3. Can the method block naturally generated harmful actions on a larger set of independent tasks, records, tools, and adaptive evasions while preserving utility?

Reviewer_rFc8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether a safety rule learned by one agent replica continues to block matching actions after later replicas take over. SGMI combines shared rule state, structured action predicates, and a trusted tool wrapper, with a conditional continuity guarantee for accepted, in-scope actions. The evaluation includes replica churn, fixed-action reuse, scope ablations, autonomous agents, and serving cost.

Critical strengths
  1. The paper makes a useful distinction between a rule being visible, matching an action, and being enforced at tool execution. The fresh-replica probes illustrate why shared memory alone cannot guarantee that a harmful action is blocked.
  2. The proposed contract specifies important preconditions, including shared rule ordering, structured predicates, and a trusted execution wrapper. This makes the intended guarantee more precise than a general claim that agents will remember safety lessons across sessions.
  3. The evaluation includes legitimate-action outcomes, a same-scope enforcement control, and audits of implementation details. Reporting these limitations gives readers a basis for judging the tradeoff between blocking harmful transfers and preserving authorized use.
Critical weaknesses
  1. The autonomous-agent study does not establish a defensive benefit. Attack success is zero for every method, including no defense, and the SGMI agent takes fallback on all 44 proposed calls without a direct structured-rule block. The fixed-action tests demonstrate enforcement in a controlled setting, but they do not show that it changes outcomes during successful autonomous attacks.
  2. The central reuse comparison changes rule scope along with enforcement. The structured banking projection omits some semantic conditions concerning subscriptions and current authorization. A broader rule can block more actions regardless of the wrapper, so this comparison does not cleanly identify how much of the effect comes from continuity enforcement.
  3. The incremental methodological contribution is limited. A same-scope shared policy control produces the same harmful and legitimate outcomes as SGMI in the principal fixed-action evaluation. The paper should explain what the proposed interface or contract establishes beyond consistently applying a shared policy at the trusted tool boundary.
  4. The evidence is narrow relative to the general agent-replica claim. Much of the controlled reuse study repeatedly applies one learned banking rule across consumers and deployments, while the rule projection and its benign overblocking are domain dependent. More independently learned rules and varied authorization conditions are needed to assess generality.
Questions for authors
  1. Can the authors evaluate autonomous episodes in which attacks succeed without defense and the agent naturally reaches an action that matches an accepted rule? What fraction of proposed calls is then directly blocked by SGMI?
  2. Can generative reuse and trusted enforcement be compared under equivalent semantic scope? Which omitted authorization conditions are essential for deciding whether the banking transfer should be blocked?
  3. What capability or empirical gain does SGMI provide beyond the same-scope shared policy control, given their identical fixed-action outcomes? Please distinguish the formal interface contribution from the enforcement result.
  4. How do the findings change across independently learned rules, varied rule specificity, and additional domains? Can the authors quantify legitimate actions rejected because a structured projection omits relevant context?
Reviewer_bJd1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies how learned safety rules can remain effective across different agent replicas. It proposes SGMI, which uses shared memory, structured rule matching, and tool level enforcement to prevent unsafe actions. The experiments show results in controlled settings.

Critical strengths
  1. The paper addresses a practical problem in agent security. Learned safety rules may be lost or ignored during later interactions.
  2. The paper separates defense failures into visibility, scope, and application, providing a clear way to understand why learned safety rules may fail.
Critical weaknesses
  1. The main experiments use only one safety record, which is insufficient to demonstrate the effectiveness of SGMI in diverse scenarios.
  2. The safety rules are manually designed for specific domains, and the harmful test actions are selected from existing benchmarks. Such experiments are not convincing enough to demonstrate the general effectiveness of SGMI.
  3. The attack success rate is 0 for all methods, even without defense(RQ4). None of the tool calls trigger SGMI's core blocking mechanism, so its effectiveness is not demonstrated.
  4. SGMI and the same-scope policy achieve identical results in all tables. The unique benefits of SGMI over the same-scope policy are not clearly demonstrated.
Questions for authors

Please see the weaknesses above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.