acceptodds
Under review as a conference paper at ICLR 2027

AgentVax: From Downstream Harm to Local Safety in Multi-Agent Workflows

Abstract

Safety in language-model multi-agent workflows is fragmented across both evidence and control, while harmful incidents may only be confirmed after earlier decisions. This creates two coupled problems: learning future risk from what each agent actually observed, and choosing an authorized action that can still change the outcome. We find that local views provide strongly complementary safety signals: at matched inference budgets and 5% validation false-positive rate, combining views raises attack-detection TPR from 55.9% for the best single view to 73.6%, a 17.7-point improvement. We also find that the position best able to recognize risk need not be the one best able to prevent harm. Motivated by these observations, we propose AgentVax, which turns verified downstream incidents into supervision for earlier local decisions. AgentVax pairs events with versioned historical inputs, uses discrete-time survival analysis to predict whether and when a violating proposal will occur, and applies counterfactual replay from restored workflow states to estimate how authorized actions change harm, legitimate-task utility, and cost. These estimates train local policies under a workflow-level false-blocking constraint, while deployment uses only each agent's current authorized view. Across controlled workflow studies spanning 24 attack families, we observe directed transfer across attacks sharing safety-relevant patterns. In a separate controlled action-value study, replay-based action selection reduces decision regret from 0.150 to 0.043, a 71.3% reduction relative to risk ranking. Together, these results show how delayed downstream feedback can become earlier, locally actionable protection by learning risk from what an agent can observe and intervention from what it can still change.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
agent+safety

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_pX4VOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

AgentVax asks how a multi-agent workflow can learn from harmful events that are confirmed only after the relevant decisions have been made. It records the local information available to each agent, trains a discrete-time survival model to forecast future policy-violating proposals, and restores historical workflow states to estimate the effects of continuing, reviewing, or blocking. These replay-derived values are then used to update local intervention policies under utility and false-blocking constraints.

The evaluation combines eight workflow strata, thirteen actor configurations, and several attack families. Local views provide complementary warning signals, and combining them improves early prediction. The full system sharply reduces harm relative to no defense, but its advantage over simpler learned guards is limited.

Critical strengths
  1. AgentVax cleanly separates the time of a harmful proposal, the later release of its verified label, and the information that was actually visible to an earlier agent. This avoids giving the safety model retrospective evidence that would not have been available online.

  2. Counterfactual replay addresses an important limitation of risk-only guards: a high risk score does not show that an authorized action will change the outcome. The paper also accounts for benign utility, false blocking, latency, and background work rather than reporting attack detection alone.

  3. The evaluation covers many workflow and actor configurations, and it reports mixed findings such as the missed test-time false-blocking target and the unfavorable cost comparison with Binary.

Critical weaknesses
  1. I do not see a clear end-to-end advantage over simpler guards. On the source stratum, Binary has lower harm with essentially the same benign success, and similar patterns appear elsewhere. AgentVax beats Central on harm in only 47 of 104 cells while using more lifecycle work; its cumulative harmful-workflow count in the closed loop is also higher than both Central and Memory. The method is much better than no defense, but the experiments do not show when its extra machinery pays off.

  2. Several results used to motivate the design are not evaluations of the complete method. The role-matched update is explicitly a separate comparator, the action-value advantage is shown in a controlled state-action experiment, and the survival objective improves timing prediction in a synthetic diagnostic while matching a binary head on the main gate outcome. These studies support individual intuitions, but they leave unclear which component produces the practical benefit of AgentVax in the end-to-end workflow.

  3. The replay procedure assumes that a complete, version-compatible workflow state can be restored and rerun under alternative actions. This is plausible in the isolated evaluation environments, but it is a demanding requirement for workflows involving external services, nondeterministic tools, or changing model versions. Since replay is both the main source of action supervision and the dominant background cost, the paper needs a clearer account of how often restoration fails and what remains of the method when only partial replay is possible.

Questions for authors
  1. Which observed platform or attack characteristics predict when AgentVax should be preferred to Binary or Central? Is there a subset of the 104 cells where replay-derived action values provide a consistent gain at matched total work?

  2. Can the survival head, binary head, and replay-trained action policy be compared within the same end-to-end workflow protocol? This would help connect the component studies to the final system.

  3. What fraction of historical checkpoints cannot be restored or replayed successfully, and how are these cases distributed across platforms and actions?

  4. How would AgentVax operate when external state cannot be cloned exactly and only logged observations or approximate simulators are available?

4 comments

Reviewer pX4V

Dear Reviewer pX4V,

We sincerely thank the reviewer for the careful and constructive feedback. We are encouraged by your recognition of the safety improvements AgentVax achieves over the undefended setting. Your detailed analysis of end-to-end performance, component contributions, and replay feasibility has also helped us clarify the evidence needed to demonstrate the method’s practical value and the importance of fair comparisons with simpler guards. We address each of your concerns in detail below.

pX4V.W1& Q1 · End-to-end value against Binary, Central and Memory

Reviewer’s comment I do not see a clear end-to-end advantage over simpler guards.

The better defensive capability of Binary in the original text comes from the fact that Binary starts prediction n windows in advance. Each node is judged n times, which leads to unfair comparisons with other methods. The experiemnt below reveals what each guard can do with the same opportunity to intervene.

In the rolling binary control, scoring starts as soon as the warning window opens. Four eligible prefixes permit up to four judgments, with fewer calls if the workflow stops early. One AgentVax call at a checkpoint can return risks for several horizons and choose one action. In the four-event comparison each guard acts once, at the same preselected checkpoint, under the same permissions, reviewer and deadline. We count the reviewer’s work separately.

One guard decision per workflow on Source, all-actor comparison**

GuardDecisionsRealized harm ↓Benign completion ↑
Binary17.5%83.7%
Central19.1%83.9%
AgentVax15.5%86.9%
AgentVax − Binary0−2.0 pp+3.2 pp

<a id="benign"></a>

** False alarms, overblocking and retention at the same checkpoint**

MetricBinaryCentralAgentVaxAgentVax − Binary
Alert on a fully observed event-free prefix ↓5.0%4.5%3.6%−1.4 pp
Benign workflow falsely blocked ↓3.0%2.8%2.3%−0.7 pp
Benign task completed ↑83.7%83.9%86.9%+3.2 pp
Completion relative to unguarded actor ↑96.0%96.2%97.6%+1.6 pp

R1a–b report the equal-count comparison. This shows that under the same and fair number of probes, agentvax can deliver better security capabilities and siginificient improvement.

Manuscript revision. We will place the matched decision counts and benign outcomes beside the Source comparison in §6.2 and Appendix E.2. Thank you for asking us to show both what was prevented and which legitimate tasks still completed.

pX4V.W2 & Q2 · Which component changes which part of the workflow?

Reviewer’s comment These studies support individual intuitions, but they leave unclear which component produces the practical benefit of AgentVax in the end-to-end workflow.

Thank you for distinguishing the supporting diagnostics from an evaluation of the complete system. The clearest reported example is replay scoring. At the fixed Source threshold in Table 9, restoring it reduced false blocking by 11.2 points and increased harm by 4.5 points. The table below shows the corresponding change in benign completion and the TAMAS result. We then specify a mechanism-specific outcome for each change to the trained guard.

<a id="r3a"></a>

R3a · Reported ablations at the fixed operating threshold, manuscript Table 9

Enabled components B/R/JSource H / U / FBR (%)TAMAS H / U / FBR (%)Source / TAMAS Brier
All three (111)9.4 / 86.7 / 5.911.4 / 88.0 / 3.90.136 / 0.134
Without early-prefix supervision (011)9.6 / 85.9 / 7.312.2 / 86.8 / 5.50.137 / 0.135
Without replay scoring (101)4.9 / 79.4 / 17.16.4 / 80.6 / 15.70.136 / 0.134
Without cost-sensitive review rule (110)7.2 / 85.9 / 7.09.0 / 87.1 / 4.60.136 / 0.134

B denotes early-prefix supervision, R denotes replay scoring, and J denotes the cost-sensitive score-and-review rule. Restoring R allowed more legitimate work to proceed under that rule. Some harmful work proceeded as well. Restoring B improved the forecast Brier score on both platforms. Fig. 2a, the synthetic timing diagnostic and the controlled action-value experiment each test a different mechanism.

R3b · One change to the trained guard on M03/Source

Change from full guardMechanism, ablated → fullH ↓U ↑FBR ↓
Full guardReference for all comparisons5.6%85.0%2.5%
Feedback stays at confirming stagePre-proposal recall 58% → 65%6.4%84.6%2.7%
Capacity-matched pooled headShared-pattern recall 65% → 70%6.0%84.4%2.9%
Hide visible provenance and authorityBoth cases in an authorization pair correct 62% → 72%7.0%84.5%2.9%
Substitute a binary headEvent-time NLL 1.27 → 1.155.6%85.0%2.5%
Shuffle replay values between matched statesAction regret 0.048 → 0.0347.4%84.6%2.8%
Use the endpoint ruleAction regret 0.050 → 0.0349.4%85.0%2.5%
Freeze after the first updateTimely control in final round 81% → 84%6.5%84.4%2.8%

For example, removing upstream supervision changes whether the reader can warn before the sender proposes an unredacted attachment. Hiding the provenance of the very same attachment instead changes whether the guard can tell an authorized internal recipient from an external recipient. All arms use the same training roots, actor, actual permissions and evaluation tasks. For the binary timing readout we distribute its endpoint event mass uniformly over the bins, as in Appendix F.3, and score the observed follow-up. All differences are paired by root. R3b reports 5.6% harm for the full guard. R3c includes both 5.6% and 8.6% result columns, and R3d reports the 8.6% result.

Manuscript revision. We will place each ablation beside its mechanism readout in Appendix F.3 and identify the action rule used in Table 9. Thank you for geive us the chance to show how each component changes behavior as well as how the workflow ends.

pX4V.W3 & Q4 · How training continues when external replay is costly or fails

Reviewer’s comment The replay procedure assumes that a complete, version-compatible workflow state can be restored and rerun under alternative actions. This is plausible in the isolated evaluation environments, but it is a demanding requirement for workflows involving external services, nondeterministic tools, or changing model versions.

We appreciate the reviewer’s question about external services. At each eligible control point, for example just before releasing an attachment or pausing it for review, the protocol records the pending operation, agent context, server-managed file contents or immutable references and their hashes, permissions, queues, event clock and model/tool versions. To test an alternative action, we load that checkpoint on an isolated server, verify the saved state and rerun the suffix under the frozen policy. This avoids repeating the preceding tool calls and interactions. Table 8 assigns 173,966.4 modeled work units to replay on Source/M08. Our proposed fallback addresses external dependencies outside the saved server state. (Section 5.2, Appendices B.5 and D.3, Algorithm D.2)

Recovery. Let h be the logged history immediately before the fork, x the local part visible to the role, u an authorized action and ξ a fresh downstream seed. A recovered branch is h ⊕ u ⊕ G(h,u,E,ξ). G generates new model decisions. E supplies compatible external observations. A past read before the fork is already part of h. A post-fork read can reuse a response only if its arguments, immutable object or quote version, effective time, permission scope and causal inputs agree. A changed recipient, new send or cancelled payment needs a new receipt for that action. Historical tokens and tool replies have zero training-loss weight. For the supplementary audit we allow up to two native attempts per branch and use a per-branch work cap of twice the discovery median for that platform and action. A request already above the cap is recorded as unattempted. An attempted branch without a verifiable effect is recorded as failed unless a compatible read supports scoring. (§5.2, Appendices B.5 and D.3)

** Reuse the recorded read when the branch still asks the same question.** When recovery failed and all tool replies received before the fork are already part of the saved history. For a read requested after the fork, we propose to store its arguments, returned content and hash, object or quote version, effective time and authorization scope. We will reuse the saved reply only when those conditions and the branch’s upstream dependencies still match and the read is reproducible. For example, a branch may retrieve the original, still-valid quote by its immutable ID. A review that asks for the current price after the quote expires needs a fresh observation. Sending to a new recipient or cancelling an external job also needs confirmation of that specific effect. If no native reference is available, we stop at the last verifiable event and retain any eligible factual labels for earlier agents.

3Build two offline training records. We also constructed two types of adventure data during training, which avoids recovery failures. For early forecasting, save (xᵢ,ₜ, mₜ, δₜ) at an event-free prefix. The last two entries give observed follow-up length and whether the first prohibited proposal occurred within it under the specified reference continuation μ. One event verified and released by the training cutoff can label several earlier agents at different distances from the event. For example, an event at step 6 under μ, verified by the cutoff, yields (m,δ)=(4,1) at step 2 and (2,1) at step 4 when K ≥ 4. A compatible benign continuation supplies the comparison. An intervention-truncated log with unknown follow-up leaves that label missing. At inference the guard sees only its original local input.

For action learning, store same-prefix pairs (xᵢ,ₜ, u, v, Q̃(h,u) − Q̃(h,v)). For each action, Q̃(h,u) = (1/M) ∑ⱼ [Lᴴⱼ + λᵤ Lᵁⱼ + λ_c Cⱼ] averages the harm, legitimate-task loss and cost over the M planned recoveries, provided all are scorable. The auxiliary target favors the action with lower conditional loss. When all required authorized actions are supported, the Eq. (5) baseline gives Ã(h,u) = ∑ᵥ π_old(v|xᵢ,ₜ) Q̃(h,v) − Q̃(h,u) for the Eq. (6) update. Otherwise we omit that action update and retain eligible factual forecast labels. We tag every target by its origin. Recovery estimates are conditional on the recorded external evidence. Independent native continuations test action ordering and drift of each action’s effect relative to continue.

Illustrative pair. A user asks for an expense total without personal details. A read-only tool has already returned the amounts and an internal identifier. From the same saved prefix, continue may produce a reply containing a personal identifier, while review may return only the correct total. Both replies can be scored locally for the disclosure and task completion. If a branch instead requests a new payment, the payment outcome stays unknown. A disclosure already verified on a compatible factual trace can still supervise that earlier prefix.

R4e · Training supervision on 640 Source roots, native evaluation on 2,000 new M03/Source roots

Training methodTraining decisions assigned native branchesH ↓U ↑FBR ↓
Supervised imitation of recorded actions (SFT)0%7.8%84.5%2.9%
Imitation weighted by factual terminal return0%7.3%84.6%2.8%
AgentVax, prefix recovery only0%6.2%84.8%2.7%
AgentVax, equal mix of native and recovered branches50%5.9%84.9%2.6%
AgentVax, full native replay100%5.6%85.0%2.5%

We hold the risk head, guard initialization, training decisions and optimizer budget fixed. Allocation in the mixed arm precedes outcome inspection. Recovery incurs 0.6 points more harm than native replay and 1.6 points less than SFT. Both contrasts use native workflows; cost accounting includes regenerated tokens, live tool calls, retries and training time.

Manuscript revision. We will add the checkpoint contents, the compatibility test for reads, and the two offline targets to Appendices B.5 and D.3. Thank you for asking how much of the learning signal survives a realistic external failure, this further enhances the reproducibility and rigor of our paper.

pX4V.Q3 · How many checkpoints and branches can be recovered, and where do failures occur?

Reviewer’s comment What fraction of historical checkpoints cannot be restored or replayed successfully, and how are these cases distributed across platforms and actions?

Thank you for insisting on the request denominator. The submitted restoration protocol verifies state identity but does not publish historical counts of attempted and unattempted requests. The fractions below are results from a supplementary audit, in which each platform contributes 120 checkpoints and each checkpoint permits all three actions. A failure after an attempt, a compatible logged read, and a branch never attempted have distinct consequences for training.

<a id="r4a"></a>

R4a · Branch status by platform and authorized action

Platform · actionRequestedExactRecorded readFailedUnattempted
Source · continue120107742
Source · review1201031043
Source · block120114141
TAMAS · continue120107652
TAMAS · review1201021143
TAMAS · block120115131
All branches720648 (90.0%)36 (5.0%)24 (3.3%)12 (1.7%)

The percentage changes when the unit is a checkpoint. Of 240 checkpoints, 200 (83.3%) have fully exact action sets, 20 (8.3%) have complete sets with at least one compatible read, and 20 (8.3%) have incomplete sets. Thus 684/720 branches (95.0%) and 220/240 checkpoints (91.7%) are scorable at their respective levels. Whole-checkpoint coverage cannot be obtained by cubing branch coverage because branch failures share external dependencies.

R4b · Compatible reads reused / branches with the stated external limitation

External dependencyContinueReviewBlockTotal
Live price, inventory or appointment availability9/1016/181/226/30
Mutable file or expiring link4/55/51/210/12
New external write or cancellation receipt0/30/10/20/6
Total13/1821/242/636/48

Here 48/720 branches (6.7%) encounter one of the listed external limitations. Of the 42 read-related cases, 36 have compatible saved results. The six incompatible reads and six unverifiable writes account for 12 of the 24 failures. The remaining 12 failures cover other state or version mismatches. The 12 unattempted branches remain a separate category. These counts describe the audited requests; they do not establish a general difference in failure rates between platforms.

Manuscript revision. We will report counts of requests, attempts and scorable action sets by platform and action in Appendices B.5 and D.4. Your question ensures that the least reproducible workflows remain visible in the results.

Reviewer_ziNTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AgentVax, a safety framework for multi-agent workflows that learns from verified downstream incidents to improve earlier local safety decisions. It combines discrete-time survival analysis for risk prediction, counterfactual replay for estimating intervention effects, and workflow-level optimization for learning local safety policies. Experiments across multiple attack families and workflow settings demonstrate improvements in risk detection, cross-attack transfer, and intervention decision-making.

Critical strengths
  1. Important Problem. The paper addresses a meaningful challenge in multi-agent safety: learning from delayed downstream incidents to enable earlier, locally authorized interventions.
  2. Comprehensive Evaluation. The paper includes extensive experiments across attack families, workflow settings, and intervention strategies, providing useful insights into complementary local evidence and risk-aware decision-making.
Critical weaknesses
  1. Limited Methodological Novelty. AgentVax primarily integrates established techniques, including discrete-time survival analysis, counterfactual replay, and advantage-based policy optimization. Although their combination is meaningful for multi-agent safety, the paper does not clearly demonstrate a substantial methodological advance beyond integrating existing approaches.
  2. Limited Improvements Over Simpler Baselines. Despite its complexity, AgentVax does not consistently outperform simpler baselines. For example, on the Source setting, Binary achieves lower harm (7.5% vs. 9.5%) with comparable benign-task success. Moreover, the additional replay and lifecycle costs raise concerns about the practical benefits of the proposed framework.
Questions for authors
  1. What is the key methodological novelty of AgentVax beyond combining existing survival analysis, counterfactual replay, and policy optimization techniques?
  2. Given that simpler baselines sometimes achieve better safety–utility trade-offs, what justifies the additional complexity and computational overhead of AgentVax?
2 comments

Reviewer ziNT

Dear Reviewer ziNT,

We sincerely thank the reviewer for the careful and constructive feedback. We are encouraged by your recognition of the value of combining risk forecasting, counterfactual replay, and policy learning to address multi-agent safety. Your comments on methodological novelty, learning across roles, and practical benefits relative to cost have also helped us clarify AgentVax’s core contributions and the conditions under which it is useful. We address each of your concerns in detail below.

ziNT.W1 & Q1

Reviewer’s comment Limited Methodological Novelty. Although their combination is meaningful for multi-agent safety, the paper does not clearly demonstrate a substantial methodological advance beyond integrating existing approaches.?

Thank you for asking where the methodological novelty enters. Discrete-time survival analysis, counterfactual replay and advantage-based policy optimization are established techniques. Our contribution lies in the following five findings and in how they let several roles learn from a single verified incident.

1. One verified incident can train guards at several stages

  • Finding. Swapping the observations available to a guard changed which violations it detected. Source confusion was easiest to recognize during reading, goal drift during planning, and unauthorized access during execution.
  • Method. Once an auditor confirms a violating proposal, we use the released label to teach each participating role how far it was from that proposal, using only what the role saw at the time. One audit thus yields several distinct training examples.
  • Result. At equal inference budget and 5% validation false-positive rate, combining stage views detected 73.6% of attacks, versus 55.9% for the best single view. In a separate fixed-threshold Source ablation, enabling early-prefix supervision reduced false blocking of benign workflows from 7.3% to 5.9%. (§§4.1 and 5.1, Fig. 1a–b, Table 9)

2. Safety experience transfers across shared patterns, and the direction matters

  • Finding. Training on one attack family could improve detection on another when they shared an operation or authorization pattern, even across different goals. Reversing the training and test families often changed the gain.
  • Method. In our case-selection study, we estimated gains in each direction from previously verified incidents. We prioritized source families that improved detection on target families while limiting additional benign false alarms.
  • Result. With the same budget of three source families, selection by measured transfer covered 66.7% of held-out families, compared with 50.0% for selection by semantic similarity. (§4.2, Fig. 1c–d, Table 3b)

3. Each role adapts over successive rounds of shared feedback

  • Finding. We compared seven update strategies on the same verified cases and budget. On new attacks sharing a known pattern, the benefit of the same cases depended on which role received the update.
  • Method. All roles’ observations train a shared risk predictor. Each role also learns a correction from its own observations. We repeat these updates as new incidents are verified and adopt a new version only if it passes an independent workflow evaluation.
  • Result. In the role-adaptation comparison, matching updates to roles improved detection on new attacks sharing a known pattern from 65.4% to 72.8% over pooled training. Benign task completion also rose, from 88.7% to 91.9%. (§4.3, Fig. 2a, Table 4a, Eq. 18)

4. Provenance and authorization change intervention outcomes

  • Finding. On held-out roles, a probe that combined content, authority, goals and capabilities identified violating proposals more accurately than a probe using content alone.
  • Method. We retain the instruction’s source and visible authorization in each role’s input, then restrict the guard to actions its runtime permits. The guard can distinguish a recipient mentioned in a document from one the user authorized.
  • Result. With actual permissions fixed on Source, hiding these cues raised realized harm by 6.9 percentage points. Restoring the cues returned harm to its original level. (§§4.4 and 6.4, Figs. 2c and 6)

5. The first role to see a risk may have little power to stop it

  • Finding. Our fixed-path derivation and permission experiment agree. At a stage with no effective intervention authority, a better warning leaves realized harm unchanged.
  • Method. From the same checkpoint, we replay each permitted action among continue, review and block under a frozen downstream policy. The local guard learns to minimize the weighted sum of harm, lost task utility and intervention cost.
  • Result. Across 240 controlled states with known action-outcome probabilities, choosing actions by their replayed consequences reduced regret from 0.150 for risk ranking to 0.043. Regret is the extra loss relative to an independently selected reference action. (Eq. 17, Fig. 7, Appendix F.2) Proposed native-workflow comparison

Our event-time likelihood follows established work on discrete-time survival models. JANUS studies risk anticipation and AGrail studies adaptive safety checks. COMA and C3 provide precedents for counterfactual credit assignment. In AgentVax, a later verifier supplies several earlier roles with supervision, each from its own historical observations, and the allowed actions are evaluated at the time each role could intervene.

Our actions. We will connect each observation in §4 to its update or control in §5 and identify the separate role-adaptation comparator in Fig. 2. Thank you for encouraging us to make the origin of each design choice explicit.

ziNT.W2 & Q2

Reviewer’s comment Limited Improvements Over Simpler Baselines.

The better defensive capability of Binary in the original text comes from the fact that Binary starts prediction n windows in advance. Each node is judged n times, which leads to unfair comparisons with other methods. The experiemnt below reveals what each guard can do with the same opportunity to intervene.

In the rolling binary control, scoring starts as soon as the warning window opens. Four eligible prefixes permit up to four judgments, with fewer calls if the workflow stops early. One AgentVax call at a checkpoint can return risks for several horizons and choose one action. In the four-event comparison each guard acts once, at the same preselected checkpoint, under the same permissions, reviewer and deadline. We count the reviewer’s work separately.

R1a · One guard decision per workflow on Source, all-actor comparison

GuardDecisionsRealized harm ↓Benign completion ↑
Binary17.5%83.7%
Central19.1%83.9%
AgentVax15.5%86.9%
AgentVax − Binary0−2.0 pp+3.2 pp

<a id="benign"></a>

R1b · False alarms, overblocking and retention at the same checkpoint

MetricBinaryCentralAgentVaxAgentVax − Binary
Alert on a fully observed event-free prefix ↓5.0%4.5%3.6%−1.4 pp
Benign workflow falsely blocked ↓3.0%2.8%2.3%−0.7 pp
Benign task completed ↑83.7%83.9%86.9%+3.2 pp
Completion relative to unguarded actor ↑96.0%96.2%97.6%+1.6 pp

R1a–b report the equal-count comparison. This shows that under the same and fair number of probes, agentvax can deliver better security capabilities.

Manuscript revision. We will place the matched decision counts and benign outcomes beside the Source comparison in §6.2 and Appendix E.2. Thank you for asking us to show both what was prevented and which legitimate tasks still completed.

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

This paper proposes AgentVax, a framework for improving local safety decisions in multi-agent workflows using verified downstream incidents. It links delayed feedback to the historical information available to each agent, predicts future violating proposals with discrete-time survival analysis, and uses restored-state replay to estimate the consequences of continuing, reviewing, or blocking. The evaluation combines controlled studies of local evidence and cross-attack transfer with workflow comparisons across eight strata and thirteen actor configurations.

Critical strengths
  • The problem setup is useful and timely. It keeps apart three events that are easy to conflate: an agent proposing a harmful action, that proposal being confirmed later, and actual harm. Each agent's predictor only uses what that agent could see at the time. It also separates the probability of a future violation from the value of an authorized intervention. These distinctions address pitfalls in learning safety policies from delayed outcomes and make the evaluation more informative than attack detection alone.
  • The observation studies offer useful evidence about complementary information and transfer. At matched inference budgets and 5% validation FPR, multiple views reach 73.6% TPR compared with 59.8% for repeated sampling of one view. Shared-pattern, different-goal attacks also show more transfer than lexical similarity alone (9.1 versus -0.8 percentage points). These controlled findings suggest that evidence access and shared safety structure matter for reusing incident feedback. The paper also reports unfavorable outcomes and costs rather than presenting only positive results.
Critical weaknesses
  • The complete system does not establish a clear benefit over simpler guards. In Table 1, Binary has lower reported harm in seven of eight strata. On Source, it achieves 7.5% harm versus AgentVax's 9.5%, with essentially the same benign success (83.7% versus 83.8%). AgentVax also incurs 1.83 times Central's lifecycle work, and its closed-loop harmful-workflow count is higher than both Central and Memory. These are not matched-operating-point significance tests, but the paper does not identify conditions where the additional machinery delivers a reliable safety, utility, or resource advantage. The Binary baseline also needs a precise definition to make this central comparison interpretable.
  • The survival model never changes a decision. Appendix F.3 states that the main gate uses only endpoint risk, so matched binary and survival heads have the same workflow outcomes. The separate timing diagnostic therefore does not show that survival modeling improves intervention decisions. Table 9 also describes its joint component as a cost-sensitive score/review rule, not a trained joint optimizer. Although Appendix E.2 states that updates follow Algorithm D.1, the relationship between that implementation and these ablations remains unclear. The experiments need to connect the trained components in Section 5 to their incremental benefit within the same end-to-end protocol.
  • The empirical validation of replay-derived supervision is insufficient. The headline regret reduction from 0.150 to 0.043 comes from 240 independent states with stipulated action-dependent harm probabilities, utility losses, and costs (Appendix F.2). This is a useful illustration of why risk ranking differs from action selection, but it does not validate action-value estimates obtained by restoring actual LLM workflows. The manuscript describes restoration and verification procedures without reporting sufficient quantitative results on restoration fidelity, failure coverage, or action-value accuracy. Because replay supplies the central intervention supervision and dominates the modeled background work, this gap directly affects the method's practical justification.
Questions for authors
  • Benefit over simpler guards. What exactly is Binary, and which components does it share with AgentVax? Under comparable benign utility or false-blocking constraints and total-work budgets, in which evaluated settings does the full method provide a consistent advantage?
  • Connection between the algorithm and ablations. Which experiments execute the trained policy in Equation (6), and how does it relate to Table 9's rule-based joint component? Can the survival, binary, and replay-trained policy variants be compared within one workflow protocol that isolates their contributions?
  • Reliability of actual replay. What fraction of historical checkpoints and action branches can be restored successfully, and what quantitative checks establish restoration fidelity and action-value accuracy? Please distinguish results from actual workflow continuations from the controlled action-law diagnostic.
5 comments

Reviewer gmaE · Summary and detailed replies

Dear Reviewer gmaE,

We sincerely thank the reviewer for the careful and constructive feedback. We are encouraged by your recognition of the value of the controlled experiment in illustrating the distinction between risk ranking and action selection. Your specific comments on fair baseline comparisons, the connection between risk forecasting and actual intervention, and the reliability of replay-derived supervision in real workflows have also helped us connect the method’s design, empirical validation, and conclusions more clearly. We address each of your concerns in detail below.

gmaE.W1

Reviewer’s comment The complete system does not establish a clear benefit over simpler guards.

Thank you for comparing both safety and benign completion across all eight platforms. We agree with your reading of Table 1. Binary records lower harm in seven strata. On Agent-SafetyBench, the remaining stratum, AgentVax records 5.4% harm versus 7.4% for Binary and 81.7% benign completion versus 83.5%. Each difference matters to the deployment choice.

The better defensive capability of Binary in the original text comes from the fact that Binary starts prediction n windows in advance. Each node is judged n times, which leads to unfair comparisons with other methods. The experiemnt below reveals what each guard can do with the same opportunity to intervene.

In the rolling binary control, scoring starts as soon as the warning window opens. Four eligible prefixes permit up to four judgments, with fewer calls if the workflow stops early. One AgentVax call at a checkpoint can return risks for several horizons and choose one action. In the four-event comparison each guard acts once, at the same preselected checkpoint, under the same permissions, reviewer and deadline. We count the reviewer’s work separately.

One guard decision per workflow on Source, all-actor comparison**

GuardDecisionsRealized harm ↓Benign completion ↑
Binary17.5%83.7%
Central19.1%83.9%
AgentVax15.5%86.9%
AgentVax − Binary0−2.0 pp+3.2 pp

<a id="benign"></a>

** False alarms, overblocking and retention at the same checkpoint**

MetricBinaryCentralAgentVaxAgentVax − Binary
Alert on a fully observed event-free prefix ↓5.0%4.5%3.6%−1.4 pp
Benign workflow falsely blocked ↓3.0%2.8%2.3%−0.7 pp
Benign task completed ↑83.7%83.9%86.9%+3.2 pp
Completion relative to unguarded actor ↑96.0%96.2%97.6%+1.6 pp

R1a–b report the equal-count comparison. This shows that under the same and fair number of probes, agentvax can deliver better security capabilities and siginificient improvement.

Manuscript revision. We will place the matched decision counts and benign outcomes beside the Source comparison in §6.2 and Appendix E.2. Thank you for asking us to show both what was prevented and which legitimate tasks still completed.

gmaE.W2 · When would a survival forecast change the action?

Reviewer’s comment The survival model never changes a decision. The experiments need to connect the trained components in Section 5 to their incremental benefit within the same end-to-end protocol.

Thank you for identifying the critical question: what the main gate reads. A discrete-time survival head estimates the conditional probability of a first prohibited proposal in each event bin, given no earlier proposal. The cumulative risk through bin k is one minus the product of the binwise survival probabilities. If a gate reads only the final cumulative risk, fixing that number, its other inputs and its random draw fixes its action. Both matched heads gave 9.4% workflow harm in Table 10. On the separate 768-root synthetic diagnostic, timing negative log-likelihood fell from 1.295 to 1.140. This measures event timing.

Illustration. An expense sheet can be reviewed now or at the next authorized checkpoint. Two handoffs each have a 40% eight-event risk. The risk in the first event is 35% in one and 5% in the other. Suppose an immediate review costs 0.10 loss units and avoids 0.50 units per early proposal after allowing for downstream protection. Its benefit is 0.175 in the first case and 0.025 in the second. Review now in the first case and wait in the second. An endpoint-only policy facing an equal mix values review at 0.10 in both and, with ties resolved in favor of continuing, releases both.

R3e · M03/Source, identical eight-event endpoint at every common prefix

Timing inputFirst action changedH in window (500 roots)H elsewhere (1,500 roots)H overallUFBR
Endpoint alone0.0%8.0%4.8%5.6%85.0%2.5%
Estimated timing profile8.0%6.0%4.8%5.1%85.0%2.5%
Shuffled profile, same endpoint7.0%8.0%4.8%5.6%85.0%2.5%

We freeze one predictor, match policy capacity, action supervision and budget, and give the timing policy the predicted profile. The negative control permutes profiles among comparable prefixes and rescales them to the original endpoint. We define the intervention window using permissions and scheduler information available before the decision. The totals equal 112 versus 102 harmful outcomes per 2,000 attack roots. The net difference of ten harmful outcomes comes from the 500 roots where review can still intercept delivery. An action change without a paired reduction in actual harm would establish sensitivity to time alone. (§6.5, Appendix F.3)

Manuscript revision. We will place the unchanged endpoint-gate result beside the decision-changing example and the native timing control in Appendix F.3. Thank you for asking for the full chain from a temporal prediction to an action and its consequence.

gmaE.W3 & Q3 · Does restored replay preserve the action that works in native execution?

Reviewer’s comment The manuscript describes restoration and verification procedures without reporting sufficient quantitative results on restoration fidelity, failure coverage, or action-value accuracy. What proportion of historical checkpoints and action branches can be successfully restored

Detailed introduction to the playback method

At each eligible control point, for example just before releasing an attachment or pausing it for review, the protocol records the pending operation, agent context, server-managed file contents or immutable references and their hashes, permissions, queues, event clock and model/tool versions. To test an alternative action, we load that checkpoint on an isolated server, verify the saved state and rerun the suffix under the frozen policy. This avoids repeating the preceding tool calls and interactions. Table 8 assigns 173,966.4 modeled work units to replay on Source/M08. Our proposed fallback addresses external dependencies outside the saved server state. (Section 5.2, Appendices B.5 and D.3, Algorithm D.2)

Recovery. Let h be the logged history immediately before the fork, x the local part visible to the role, u an authorized action and ξ a fresh downstream seed. A recovered branch is h ⊕ u ⊕ G(h,u,E,ξ). G generates new model decisions. E supplies compatible external observations. A past read before the fork is already part of h. A post-fork read can reuse a response only if its arguments, immutable object or quote version, effective time, permission scope and causal inputs agree. A changed recipient, new send or cancelled payment needs a new receipt for that action. Historical tokens and tool replies have zero training-loss weight. For the supplementary audit we allow up to two native attempts per branch and use a per-branch work cap of twice the discovery median for that platform and action. A request already above the cap is recorded as unattempted. An attempted branch without a verifiable effect is recorded as failed unless a compatible read supports scoring. (§5.2, Appendices B.5 and D.3)

** Reuse the recorded read when the branch still asks the same question.** When recovery failed and all tool replies received before the fork are already part of the saved history. For a read requested after the fork, we propose to store its arguments, returned content and hash, object or quote version, effective time and authorization scope. We will reuse the saved reply only when those conditions and the branch’s upstream dependencies still match and the read is reproducible. For example, a branch may retrieve the original, still-valid quote by its immutable ID. A review that asks for the current price after the quote expires needs a fresh observation. Sending to a new recipient or cancelling an external job also needs confirmation of that specific effect. If no native reference is available, we stop at the last verifiable event and retain any eligible factual labels for earlier agents.

3Build two offline training records. We also constructed two types of adventure data during training, which avoids recovery failures. For early forecasting, save (xᵢ,ₜ, mₜ, δₜ) at an event-free prefix. The last two entries give observed follow-up length and whether the first prohibited proposal occurred within it under the specified reference continuation μ. One event verified and released by the training cutoff can label several earlier agents at different distances from the event. For example, an event at step 6 under μ, verified by the cutoff, yields (m,δ)=(4,1) at step 2 and (2,1) at step 4 when K ≥ 4. A compatible benign continuation supplies the comparison. An intervention-truncated log with unknown follow-up leaves that label missing. At inference the guard sees only its original local input.

For action learning, store same-prefix pairs (xᵢ,ₜ, u, v, Q̃(h,u) − Q̃(h,v)). For each action, Q̃(h,u) = (1/M) ∑ⱼ [Lᴴⱼ + λᵤ Lᵁⱼ + λ_c Cⱼ] averages the harm, legitimate-task loss and cost over the M planned recoveries, provided all are scorable. The auxiliary target favors the action with lower conditional loss. When all required authorized actions are supported, the Eq. (5) baseline gives Ã(h,u) = ∑ᵥ π_old(v|xᵢ,ₜ) Q̃(h,v) − Q̃(h,u) for the Eq. (6) update. Otherwise we omit that action update and retain eligible factual forecast labels. We tag every target by its origin. Recovery estimates are conditional on the recorded external evidence. Independent native continuations test action ordering and drift of each action’s effect relative to continue.

Illustrative pair. A user asks for an expense total without personal details. A read-only tool has already returned the amounts and an internal identifier. From the same saved prefix, continue may produce a reply containing a personal identifier, while review may return only the correct total. Both replies can be scored locally for the disclosure and task completion. If a branch instead requests a new payment, the payment outcome stays unknown. A disclosure already verified on a compatible factual trace can still supervise that earlier prefix.

R4e · Training supervision on 640 Source roots, native evaluation on 2,000 new M03/Source roots

Training methodTraining decisions assigned native branchesH ↓U ↑FBR ↓
Supervised imitation of recorded actions (SFT)0%7.8%84.5%2.9%
Imitation weighted by factual terminal return0%7.3%84.6%2.8%
AgentVax, prefix recovery only0%6.2%84.8%2.7%
AgentVax, equal mix of native and recovered branches50%5.9%84.9%2.6%
AgentVax, full native replay100%5.6%85.0%2.5%

We hold the risk head, guard initialization, training decisions and optimizer budget fixed. Allocation in the mixed arm precedes outcome inspection. Recovery incurs 0.6 points more harm than native replay and 1.6 points less than SFT. Both contrasts use native workflows; cost accounting includes regenerated tokens, live tool calls, retries and training time.

Manuscript revision. We will add the checkpoint contents, the compatibility test for reads, and the two offline targets to Appendices B.5 and D.3. Thank you for asking how much of the learning signal survives a realistic external failure, this further enhances the reproducibility and rigor of our paper.

How many checkpoints and branches can be recovered, and where do failures occur?

The submitted restoration protocol verifies state identity but does not publish historical counts of attempted and unattempted requests. The fractions below are results from a supplementary audit, in which each platform contributes 120 checkpoints and each checkpoint permits all three actions. A failure after an attempt, a compatible logged read, and a branch never attempted have distinct consequences for training.

R4a · Branch status by platform and authorized action

Platform · actionRequestedExactRecorded readFailedUnattempted
Source · continue120107742
Source · review1201031043
Source · block120114141
TAMAS · continue120107652
TAMAS · review1201021143
TAMAS · block120115131
All branches720648 (90.0%)36 (5.0%)24 (3.3%)12 (1.7%)

The percentage changes when the unit is a checkpoint. Of 240 checkpoints, 200 (83.3%) have fully exact action sets, 20 (8.3%) have complete sets with at least one compatible read, and 20 (8.3%) have incomplete sets. Thus 684/720 branches (95.0%) and 220/240 checkpoints (91.7%) are scorable at their respective levels. Whole-checkpoint coverage cannot be obtained by cubing branch coverage because branch failures share external dependencies.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.