- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
ORBIT learns intervention policies for tool-using agents using evidence accumulated across actions and episodes. It combines Bayesian beliefs with an offline-fitted policy, updates beliefs online, and models monitor incentives through obedience constraints on report deviations. Branch replay evaluates allow, inspect, and block decisions; AgentDojo evaluates allow and campaign-ending stop decisions. The paper provides conditional theoretical guarantees, comparisons with threshold and direct policies, memory ablations, adaptive attacks, and modeled monitor-response experiments. The results support inspection benefits and memory-dependent prevention gains, with important costs in denied legitimate work.
Critical strengths
S1. Meaningful evidence for inspection at comparable benign cost.
In Table 5, ORBIT-postrule achieves 37.1% early stopping at 19.35 weighted benign interruptions per 1,000 original benign decisions, compared with 18.5% at 20.71 for DBN thresholding. The expanded six-seed experiment also has favorable mean early stopping and benign cost (Table 14). These results provide evidence beyond simply interrupting more activity. The targeted inspection-evidence concealment study usefully identifies conditions under which this benefit deteriorates.
S2. Useful controls and careful distinctions between guarantees and measurements.
The evaluation includes full-memory Direct and threshold controls, multiple calibration allowances, paired uncertainty, and adaptive attacks reexecuted with task-checker-confirmed harm. Keeping canceled legitimate work in the cost denominator is appropriate. The η=0.1 audit establishes that obedience can exclude otherwise realizable action policies, while the authors correctly refrain from claiming a reliable performance improvement. The theoretical guarantees state their assumptions, and the manuscript acknowledges limitations of the recording-based coverage diagnostic.
Critical weaknesses
W1. The testbed construction and monitor-information assumptions limit the scope and reproducibility of the empirical conclusions.
Appendix H.1 refers to accepted cached replay corpora without sufficiently describing their generation and acceptance. The released evaluator requires node AUC ≤ .60, single-edge AUC ≤ .70, and connected-model AUC between .80 and .95. The node and connected checks score the corpus test partition. These checks enforce a weak-local/strong-connected-evidence regime, relevant to a method that accumulates evidence. The mechanism label prioritizes active_deviation, and the connected model uses DBN machinery, so these criteria alone do not determine ORBIT's advantage over DBN. Appendix J also identifies cached recordings and fitted artifacts that are required for reproduction but omitted from the source release.
In AgentDojo, the released runner obtains the permission monitor's allowed tools from reference-solution tool usage and matches taint against actual injection text. Deployment may provide different information. The existing both-channels-disabled ablations retain some memory gains and are useful evidence for scoping this assumption.
Q1 — Clarification of the existing protocol: Please describe corpus generation, when acceptance criteria were fixed, and how acceptance was separated from final policy evaluation; provide the frozen inputs or an auditable reconstruction procedure. Please also state the benchmark-derived monitor inputs and the deployment scope supported by the existing ablations.
W2. Persistent memory's benefit needs clearer characterization at comparable achieved legitimate-work cost.
At 5% observation without taint, prevention increases from 13.33% with episode reset to 51.33% with full memory, while legitimate work denied rises from 11.55% to 45.01% (Table 2). At adaptive search budget 16, harmful completion falls from 18.50% to 7.00%, with denial increasing from 10.99% to 43.96% (Table 7). The authors acknowledge these costs and provide Direct controls and tighter allowances. However, a common calibration allowance does not establish a favorable comparison at the same achieved cost; at the 20% allowance, memory's mean prevention advantage over episode reset appears in only three of six settings, with all corresponding paired intervals including zero.
Q2 — Additional analysis relevant to the memory claim: Please compare the existing ORBIT, episode-reset, Direct, and threshold policy families over overlapping achieved-cost ranges using operating points selected without test feedback, or limit the memory claim to the demonstrated operating points. The adaptive result remains a useful overall-system comparison; campaign stops also reduce attacker feedback, as the paper states. This concern is specific to memory, while branch replay already provides positive inspection results at approximately comparable benign cost.
Overall assessment. I lean toward weak rejection because the testbed and information assumptions need clearer substantiation, and the persistent-memory evidence does not yet establish its favorable achieved-cost regime sufficiently. The inspection result and experimental controls are substantive strengths. Clearer construction provenance, appropriately scoped deployment claims, and the targeted trade-off analysis could materially change this borderline assessment.
Questions for authors
See weaknesses
We thank the reviewer for the careful assessment and for recognizing the inspection results, experimental controls, and distinction between theoretical guarantees and empirical findings.
W1 (testbed construction): yes the replay corpus is conditioned on a weak-local/strong-connected-evidence regime but corpus acceptance and intervention-policy selection are distinct. Policies are trained on training data and selected using validation or calibration data, without using test outcomes for policy selection. However, the evaluator's use of test-partition AUC for corpus acceptance limits claims of unconditional testbed generalization. We acknowledge this distinction. In AgentDojo, permission and taint monitoring use benchmark-derived information; the ablations removing both channels provide evidence under restricted monitoring, but do not establish deployment-wide generalization. Appendix J also acknowledges the reproducibility limitations of the current release.
W2 (memory and benign cost): also agree that equal calibration limits do not imply equal achieved interruption costs. Our claim is that persistent memory improves prevention at the reported operating points, not that it universally dominates episode-reset policies at equal benign cost. Section 3.4, Tables 2 and 11, and Appendix F explicitly document the associated disruption and weaker results under tighter limits.
We appreciate the reviewer's recognition that the inspection results demonstrate meaningful benefits at comparable benign cost, independently of the persistent-memory trade-off.