acceptodds
Under review as a conference paper at ICLR 2027

Holding Agents Accountable with Verifiable Agent Insurance

Abstract

While AI agents are useful for many tasks, they regularly misbehave in destructive ways, such as deleting important data. While prior work has focused on preventing such misbehavior, we consider how to recover after an agent misbehaves. We propose verifiable agent insurance (VAI), a framework for agent providers to commit financially to promises, a list of programmatically-specified policies that an agent is meant to follow. These promises are embedded in a smart contract on an immutable ledger (e.g., a blockchain) to prevent tampering, along with a monetary escrow. A user who detects a violation of committed promises can prompt a verifier to review agent execution traces, which are logged through the agent harness's shared tool-execution path, without instrumenting each tool separately. Upon verification of the violation, the agent provider's escrow funds are paid to the reporting user. We instantiate VAI as an end-to-end implementation on an Ethereum layer-2 network, and evaluate its coverage on three prominent agent benchmarks, two of which we re-annotated to obtain task-level policy-compliance labels. Currently, many agents do not emit user authorizations and security-relevant information as explicit, structured data objects—rather, they embed the data needed to verify policy violations in natural language exchanges. Even so, VAI manages to detect violations in 42.2% of runs containing violations across three benchmarks. By augmenting agents to explicitly emit such objects, violation detection rises to 83.0% of traces. Our findings suggest that for improved auditability, agents should be co-designed with their security policies to emit explicit, structured records of user authorization and security-relevant actions.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI agents, agent accountability, AI safety, financial guarantees, smart contracts

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_imYqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes verifiable agent insurance, a framework linking programmatically specified behavioral promises to committed execution traces and provider-funded escrow. Users can challenge suspected violations, and a trusted verifier checks the evidence before settlement. The authors implement the system with two agent harnesses and deploy an escrow contract on Base. Benchmark experiments measure run-level violation detection and examine how additional structured records could expand coverage.

Critical strengths
  1. The paper asks a worthwhile question about what happens after prevention fails. Connecting explicit behavioral commitments to evidence and financial recourse gives agent accountability a concrete operational meaning.
  2. The contribution includes a working implementation rather than only a conceptual proposal. Harness integrations, trace commitments, settlement rules, mainnet tests, and runtime measurements make the design easier to assess.
  3. The empirical analysis identifies an actionable limitation of current agents: information needed for auditing often remains embedded in natural language. The discussion of consent, permission scope, and structured records is useful beyond the proposed insurance mechanism.
Critical weaknesses

The 83.0% coverage result needs more careful qualification. The measured detection rate on the original benchmark traces is 42.2%. Appendix D.3 estimates additional coverage through retrospective review and supports proposed checks with 183 constructed examples. This is useful evidence, but it differs from instrumenting the agents and rerunning the benchmarks. Some statements in the abstract and main results read as though the latter experiment was performed. The paper should consistently distinguish measured detection from potential coverage.

The annotation validation is not sufficiently independent. Most policy-compliance labels come from a three-LLM committee. Its agreement with the five-labeler majority is not an independent accuracy check because those same three LLMs also contribute to that majority. The calibration samples are small, and Appendix D.4 shows that reasonable interpretations can disagree about whether a violation occurred. A separate human-adjudicated sample would strengthen confidence in both the coverage estimates and the reported 5.8% false-positive rate, especially when verdicts trigger payments.

A valid claim does not necessarily receive compensation. Appendix B.2 states that filing a claim does not reserve its payout, other claims may exhaust the balance, and new claims are rejected when reserves are insufficient. This substantially limits the accountability guarantee: a recorded and verified violation may remain unpaid. The limitation is disclosed, but should be reflected more prominently in the framework’s claims. Concurrent-claim and reserve-exhaustion tests would help establish what protection a registered user actually receives.

Questions for authors
  1. Can the added-record coverage estimate be validated on a representative subset using genuinely instrumented agent executions, including the provenance of authorization records?
  2. How do the coverage and false-positive estimates change on a separate, human-adjudicated sample that excludes the calibration cases?
  3. What guarantee does a registered session receive when several valid claims compete for insufficient reserves? Could coverage admission or payout reservation provide a stronger guarantee?
3 comments
AuthorsOpenReview ↗

Response to Reviewer imYq (1/2)

Thank you for your constructive suggestions. We address your comments below:

Weakness 1

“The 83.0% coverage result needs more careful qualification. The measured detection rate on the original benchmark traces is 42.2%. Appendix D.3 estimates additional coverage through retrospective review and supports proposed checks with 183 constructed examples. This is useful evidence, but it differs from instrumenting the agents and rerunning the benchmarks. Some statements in the abstract and main results read as though the latter experiment was performed. The paper should consistently distinguish measured detection from potential coverage.”

Good point, we have replaced the projection with actual measurements. For each benchmark, we listed the rules (policies) the agent is supposed to follow. If VAI could not check a rule because the harness does not log the information the rule needs, we manually evaluated whether a harness could record that information as structured data, without interpreting natural language. If so, we updated the harness and registered a promise for that rule. In Tau-Bench, the only harness update consisted of one tool that lets the agent ask the user to approve an exact tool call before making it. The user answers ‘approve’ or ‘deny’ and the harness records the request and the answer. In ClawsBench, the only update was that the harness now logs what each tool call returned to the agent, so a promise can check, for example, that the agent read a document before sharing it outside the organization. We then ran both benchmarks again and VAI detected 79.8% of the total violations detected with the updated harness. It doesn't match the previous projected 83% exactly because running the benchmarks again generated a new corpus of agent-user interactions. We have updated the paper in the abstract, Section 1, Section 5.1, Table 1 and Section 5.2.

Weakness 2 / Question 1

“The annotation validation is not sufficiently independent. Most policy-compliance labels come from a three-LLM committee. Its agreement with the five-labeler majority is not an independent accuracy check because those same three LLMs also contribute to that majority. … Can the added-record coverage estimate be validated on a representative subset using genuinely instrumented agent executions, including the provenance of authorization records?”

We agree and have since removed that Dawid-Skene study and the five-labeler comparison. Three LLMs now label every run independently and we take their 2-of-3 majority as the label. We then sample 30 random runs per benchmark and have two human annotators independently label them, without seeing each other's labels or the LLM labels. Neither annotator had seen the traces before doing this annotation. Before resolution, the two annotators agree on 26/30 runs (Cohen’s kappa = 0.64) on Tau-Bench and 27/30 on ClawsBench (Cohen’s kappa = 0.52). The human annotators then discuss and resolve their disagreements and establish a human baseline. After discussion, the human annotators agreed on all annotations. We compare the LLM majority on those same 60 runs with the human baseline and they agree on 27/30 Tau-Bench runs (Cohen’s kappa = 0.71) and 24/30 ClawsBench runs (Cohen’s kappa = 0.14). The low Cohen’s kappa on ClawsBench is because violations are rare in ClawsBench and the human baseline labeled 5 violations and the LLM majority found only 1 of them. Section 5.1 and Appendix D.2 (Table 7) have been updated to reflect this.

AuthorsOpenReview ↗

Response to Reviewer imYq (2/2)

Weakness 3

“A valid claim does not necessarily receive compensation. Appendix B.2 states that filing a claim does not reserve its payout, other claims may exhaust the balance, and new claims are rejected when reserves are insufficient. This substantially limits the accountability guarantee: a recorded and verified violation may remain unpaid. The limitation is disclosed, but should be reflected more prominently in the framework's claims. Concurrent-claim and reserve-exhaustion tests would help establish what protection a registered user actually receives.”

Thank you for noticing this. We have updated VAI so that when a challenge is filed, it sets aside the payout p for that challenge, and if R < p at filing time then we would no longer reject that challenge but instead queue the challenges sorted by the filing time. If a queued challenge wins, the user gets the bond back at once and the payout once funds reach it. While nothing forces the provider to add funds and a proven claim can stay owed, the reserve is public and we deliberately leave the price discovery and potential restaking for future work. Our contract tests cover concurrent claims and reserve-exhaustion: claims are paid in filing order whatever the order of the verdicts, a provider's own claim cannot take the payout of a claim filed before it, and a claim filed against an empty reserve is paid once a deposit arrives. Implementation details can be found in Section 3.1, Section 3.3, Algorithm 1 (Appendix A), Appendix B.2 and Section 7.

Question 2

“How do the coverage and false-positive estimates change on a separate, human-adjudicated sample that excludes the calibration cases?”

Now using the sample of the 60 tasks that’s randomly sampled from the benchmarks mentioned above, we score VAI with the updated harness twice, once against the resolved human labels, and once against the LLM majority labels. “Detected” is the number of runs VAI detects out of the runs that were labeled as violations; “False alarms” is the number of runs VAI flags out of the runs that were labeled as no violations. We present the numbers below, in the format (# Detected / # Annotated Violations), where the annotated violations are taken from the human annotators (left column) and the LLM majority (right). These results are reported only for the 60-task validation set on which we ran a human evaluation.

Against the human baselineAgainst the LLM labels
τ³-bench detected13/2513/22
τ³-bench false alarms0/50/8
ClawsBench detected2/51/3
ClawsBench false alarms1/252/27

Question 3

“What guarantee does a registered session receive when several valid claims compete for insufficient reserves? Could coverage admission or payout reservation provide a stronger guarantee?”

We added a mechanism for reserving funds at the time of submitting a claim, so now the payout is set aside at filing and claims are funded in filing order. We did not adopt coverage admission since a user can open infinite sessions, so the reserved amount needed for each promise would be unbounded. This is described in Section 3.1 and Appendix B.2.

Thank you again for your careful review and thoughtful feedback. We would like to follow up before today's discussion deadline to see whether our response has addressed your concerns. Please let us know if there is anything we can further clarify.

Reviewer_H5wjOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper introduces VAI, a cryptographic accountability and financial recourse framework for AI agents. VAI binds programmatically defined policy promises to an on-chain escrow smart contract on Base mainnet (Ethereum Layer-2), instruments agent harnesses at shared tool-dispatch hooks to log actions off-chain, periodically commits cryptographic checkpoint hashes to the ledger, and enables post-hoc verification and automated slashing upon reported violations. Evaluating across AgentDojo, τ3\tau^3-bench, and ClawsBench, VAI achieves a 42.2% baseline TPR at sub-cent transaction costs, while positing that structured telemetry co-design could theoretically raise coverage to 83.0%.

Critical strengths
  • Establishes an end-to-end framework combining lightweight tool-dispatch hooks, periodic trace hashing, and automated smart-contract escrow slashing on Base mainnet with minimal runtime overhead.
  • Effectively demonstrates that current agent architectures bury security-critical state and user consent in free-text conversation history, identifying telemetry co-design as a key requirement for post-hoc auditability.
  • Annotates and releases 388 fine-grained, task-level policy-compliance labels for τ3\tau^3-bench and ClawsBench, providing valuable ground truth for intermediate procedural compliance which can be useful benchmark.
Critical weaknesses
  • The protocol suffers from a fundamental Time-of-Action to Time-of-Commitment window: because tool calls execute locally before on-chain hashing, a rational untrusted provider can alter or drop incriminating records prior to checkpointing. Furthermore, pooling an unearmarked reserve (RtR_t) while paying out user-induced prompt injections introduces severe Sybil payout-farming vulnerabilities, and the headline 83.0% TPR relies on post-hoc theoretical LLM judging rather than live end-to-end benchmark execution.
  • Because tool calls execute and write to a provider-managed local SQLite store before batched on-chain checkpoints are committed (every kk calls or rr seconds), a self-interested provider can alter or truncate incriminating terminal records prior to checkpointing. How does VAI prevent a dishonest provider from truncating the local execution trace prior to the next batched checkpoint when an irreversible violation occurs?
  • Compensating user-induced violations (such as deliberate prompt injections) from an unearmarked shared reserve pool (RtR_t) exposes the protocol to Sybil payout farming and provider front-running/self-slashing, which can drain funds and deny recourse to legitimate users. How can the protocol prevent economic arbitrage loops where users deliberately prompt-inject agents to repeatedly drain fixed bounty payouts (pp) from the shared reserve (RtR_t)?
  • The headline 83.0% TPR claim is an LLM-judged theoretical checkability estimate tested on synthetic unit tests rather than live benchmark runs with modified harnesses. Furthermore, 99.4% of evaluated violation runs (AgentDojo and τ3\tau^3-bench) were conducted offline on static logs without testing the integrated Hermes or OpenClaw adapters. Does the 83.0% coverage claim hold when running live agents with structured logging, and what is the resulting false-positive impact on compliant sessions?
  • Shared dispatch hooks inspect only top-level tool names and inputs; when agents invoke general-purpose shell (bash) or code execution tools, internal file deletions, process executions, or network connections bypass parameter-level predicate checks.
Questions for authors

Please see weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer H5wj (1/2)

Thank you for your careful reading and thoughtful suggestions regarding the design of our system, we appreciate the feedback. We address your comments below:

Weakness 1

“Because tool calls execute and write to a provider-managed local SQLite store before batched on-chain checkpoints are committed (every k calls or r seconds), a self-interested provider can alter or truncate incriminating terminal records prior to checkpointing. How does VAI prevent a dishonest provider from truncating the local execution trace prior to the next batched checkpoint when an irreversible violation occurs?”

Good point, thank you for raising this. We have updated VAI to prevent the agents from modifying records so that it generates a key pair for each session and commits the public key on chain. Using that key, the SDK signs a checkpoint before each tool call and runs the call only after the user's device has verified and stored the signature. After each tool call, the SDK signs the traces so far and sends the signature to the user device again. The user only needs to keep the latest signature and would submit it when filing a challenge so that any alteration to records at any time would result in a different signature and would be ruled against the provider. We tested this on Base mainnet and the specific implementation and signature parameters can be found in Section 3.2, Section 3.3, Appendix C.2, Algorithms 1–2 and Appendix D.4. On average, this addition adds 49ms to each tool call with OpenClaw and 62ms with Hermes (Table 8). For comparison, the OpenClaw agent in ClawsBench takes 3.8 seconds per tool call on average, including model generation, so the addition is about 1.3% of the time per tool call and presents no significant latency bottleneck.

Weakness 2

“Compensating user-induced violations (such as deliberate prompt injections) from an unearmarked shared reserve pool (R) exposes the protocol to Sybil payout farming and provider front-running/self-slashing, which can drain funds and deny recourse to legitimate users. How can the protocol prevent economic arbitrage loops where users deliberately prompt-inject agents to repeatedly drain fixed bounty payouts (p) from the shared reserve (R)?”

Paying for induced violations is the intended design. If a user can prompt-inject an agent to cause it to violate its own policies, then the agent by definition does not reliably follow those policies; we argue that the agent provider in this case should be slashed. Moreover, although the provider could indeed drain its own reserve, the reserve and challenges are public on-chain, so users can see that a given promise is no longer funded before using the agent. This is the most we can hope to provide a user. We have also updated VAI so that when a challenge is filed, it sets aside the payout p for that challenge, and if R < p at filing time then we would queue the challenges sorted by the filing time. This prevents scenarios where the provider drains the pool over the course of a single claim. Implementation details can be found in Section 3.1, Section 3.3, Algorithm 1 (Appendix A), Appendix B.2 and Section 7. While we cannot fully prevent provider front-running concurrently with a user using the agent (but before filing a claim), we envision the agent provider ecosystem being eventually influenced in part by reputation systems (versions of this are already happening, such as in ERC-8004) and possibly even by regulation. Hence, if an agent's reserves are consistently depleted, there may be social mechanisms by which users can learn this and avoid a given provider. However, we expect that fully solving front-running may not have an easy technical solution, as mirrored in the decentralized finance ecosystem at large.

AuthorsOpenReview ↗

Response to Reviewer H5wj (2/2)

Weakness 3

“The headline 83.0% TPR claim is an LLM-judged theoretical checkability estimate tested on synthetic unit tests rather than live benchmark runs with modified harnesses… Does the 83.0% coverage claim hold when running live agents with structured logging, and what is the resulting false-positive impact on compliant sessions?”

Good point, we have replaced the projection with actual measurements on Tau-Bench and ClawsBench, which corroborate our earlier projections within a small margin of error. For each benchmark, we listed the rules (policies) the agent is supposed to follow. If VAI could not check a rule because the harness does not log the information the rule needs, we manually evaluated whether a harness could record that information as structured data, without interpreting natural language. If so, we updated the harness and registered a promise for that rule. In Tau-Bench, the only harness update consisted of one tool that lets the agent ask the user to approve an exact tool call before making it. The user answers ‘approve’ or ‘deny’ and the harness records the request and the answer. In ClawsBench, the only update was that the harness now logs what each tool call returned to the agent, so a promise can check, for example, that the agent read a document before sharing it outside the organization. We then ran both benchmarks again and VAI detected 79.8% of the total violations detected with the updated harness. It doesn't match the previous projected 83% exactly because running the benchmarks again generated a new corpus of agent-user interactions. We have updated the paper in the abstract, Section 1, Section 5.1, Table 1 and Section 5.2.

Weakness 4

“Shared dispatch hooks inspect only top-level tool names and inputs; when agents invoke general-purpose shell (bash) or code execution tools, internal file deletions, process executions, or network connections bypass parameter-level predicate checks.”

Yes this is a limitation of VAI. We do not view VAI as a catch-all solution, but as an early step in the problem of agent auditing and recourse. Your observation this could be mitigated through making the tool calls more structural, such as defining a tool that deletes a file instead of having the agent running rm -rf in a shell. We now state this in Section 7. However, we expect that designing a fully expressive system in this way would increase overhead prohibitively, both for the user (interface requirements) and for developers; hence, this may not be practical.

Thank you again for your careful review and thoughtful feedback. We would like to follow up before today's discussion deadline to see whether our response has addressed your concerns. Please let us know if there is anything we can further clarify.

Reviewer_TwNgOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

This paper studies post-hoc accountability and financial recourse after AI agents misbehave. It proposes Verifiable Agent Insurance (VAI), a framework that combines programmatically specified policies ("promises"), execution trace logging, and on-chain escrow to compensate users for objective policy violations. Under VAI, agent providers deposit collateral into a smart contract on an immutable ledger; the agent runtime harness records tool calls and authorization events through shared dispatch hooks while committing periodic checkpoint hashes and a final trace hash on-chain; and when a user files a claim with a bond, a trusted independent verifier checks trace integrity against on-chain commitments and evaluates policy predicates, triggering automated smart-contract settlement.

The authors integrate VAI into two open-source agent harnesses (Hermes and OpenClaw) and validate end-to-end settlement on Base mainnet across 14 challenges in 11 sessions. Across three benchmarks (AgentDojo, τ3\tau^3-bench, and ClawsBench, with the latter two re-annotated for task-level policy compliance), VAI detects at least one violation in 42.2% (294/696) of violating runs using existing execution records, with a pooled false-positive rate of 5.8% (18/313). Based on a missing-information analysis of undetected violations, the authors further estimate that augmenting agents to emit additional structured records could increase violation coverage to 83.0% (578/696).

Critical strengths
  1. Practical Problem Formulation and End-to-End System Prototype: Because pre-execution safeguards cannot prevent all agent failures, establishing executable post-hoc accountability is a practically important direction. The paper presents a complete prototype that links programmatic policy predicates, cryptographic trace integrity, independent verification, and automated escrow payouts. By providing concrete integrations with mainstream harnesses (Hermes and OpenClaw), live mainnet testing on Base, and latency/fee measurements, the work turns post-hoc agent liability into a concrete, empirically inspectable systems problem.

  2. Valuable Diagnostic Insights into Agent Auditability: The paper offers a clear and useful analysis of why many policy violations currently evade programmatic detection: critical context—such as user consent, dynamic permission changes, and eligibility checks—often resides only in unstructured natural-language exchanges rather than structured execution records. Analyzing these missing records and formulating corresponding verification checks provides actionable guidance for harness developers on what structured telemetry must be exposed to co-design auditable agents.

Critical weaknesses
  1. Headline Coverage Claim (83.0%) Is an Analytical Estimate Rather Than an Empirical Evaluation on Augmented Runs: A central quantitative claim of the paper—that by augmenting agents to emit structured objects, "violation detection rises to 83.0% of traces" (Abstract) and "the overall TPR jumps to 83%" (Section 5.2)—is significantly overstated relative to the actual evidence. In Table 1, only the 42.2% rate reflects actual detection on benchmark traces (with τ3\tau^3-bench at 13.5%–20.1% and ClawsBench at 0.0%), whereas the 83.0% figure counts runs classified by an LLM-as-a-judge and human review as "Could detect with added records." While the 51 implemented checks and 183 constructed examples in Appendix D.3 confirm that the predicate logic functions on synthetic SDK inputs, they do not demonstrate that these records can be reliably produced and verified during live benchmark execution. Crucially, capturing information such as user authorization or dynamic scope updates (which also account for 13/18 false positives when expressed in free text, e.g., supplying a spouse's email in natural language) often requires non-trivial UI/interaction redesigns or semantic interpretation, rather than merely adding logging hooks to existing tool calls.

  2. Lack of Independent Human Validation for Reference Labels: The ground-truth violation labels used to evaluate VAI on τ3\tau^3-bench and ClawsBench lack rigorous independent validation. Aside from two 30-run calibration samples, the remaining 298 τ3\tau^3-bench runs and 30 ClawsBench runs are annotated solely by a Dawid–Skene weighted committee of three LLMs. Crucially, the high agreement reported on the calibration sets (29/30 and 30/30) is measured between this three-LLM committee and the five-labeler majority (or Dawid–Skene consensus) formed by the exact same three LLMs plus two human annotators. Because the three LLMs already constitute a 3-of-5 majority in the combined pool, this comparison is partially circular and does not measure accuracy against an independent human reference. Given that VAI verdicts are intended to trigger real financial payouts, the absence of an independent human audit undermines confidence in the reference labels when interpreting both missed violations and false positives.

Questions for authors

For the violations categorized under "Could detect with added records" (which bridge the gap from 42.2% to 83.0%), could you systematically break down which required records can be deterministically extracted from existing tool calls/results versus which ones require new user-interaction mechanisms or trusted semantic parsing of natural language? Furthermore, can you actually instrument these augmented records on a representative subset of benchmark tasks and report both the resulting empirical detection rate and the false-positive rate?

4 comments
AuthorsOpenReview ↗

Response to Reviewer TwNg

Thank you very much for your constructive suggestions. We address your comments and questions below:

Weakness 1 / Question 1

“Headline Coverage Claim (83.0%) Is an Analytical Estimate Rather Than an Empirical Evaluation on Augmented Runs. For the violations categorized under "Could detect with added records"... could you systematically break down which required records can be deterministically extracted from existing tool calls/results versus which ones require new user-interaction mechanisms? Can you instrument these records on a representative subset of benchmark tasks and report both the resulting empirical detection rate and the false-positive rate?”

Good point, we have replaced the projection with actual measurements on Tau-Bench and ClawsBench, which corroborate our earlier projections within a small margin of error. For each benchmark, we listed the rules (policies) the agent is supposed to follow. If VAI could not check a rule because the harness does not log the information the rule needs, we manually evaluated whether a harness could record that information as structured data, without interpreting natural language. If so, we updated the harness and registered a promise for that rule. In Tau-Bench, the only harness update consisted of one tool that lets the agent ask the user to approve an exact tool call before making it. The user answers ‘approve’ or ‘deny’ and the harness records the request and the answer. In ClawsBench, the only update was that the harness now logs what each tool call returned to the agent, so a promise can check, for example, that the agent read a document before sharing it outside the organization. We then ran both benchmarks again and VAI detected 79.8% of the total violations detected with the updated harness. It doesn't match the previous projected 83% exactly because running the benchmarks again generated a new corpus of agent-user interactions. We have updated the paper in the abstract, Section 1, Section 5.1, Table 1 and Section 5.2.

Weakness 1 (continued)

“Crucially, capturing information such as user authorization or dynamic scope updates (which also account for 13/18 false positives when expressed in free text, e.g., supplying a spouse's email in natural language) often requires non-trivial UI/interaction redesigns or semantic interpretation, rather than merely adding logging hooks to existing tool calls.”

For Tau-Bench and ClawsBench our updates were minimal and required no extensive UI/interaction redesigns. For instance, our sole update to the harness in Tau-Bench is one approval tool that the agent can invoke to ask for user's permission where the user can select yes or no. This practice is common in agent harnesses such as OpenClaw, Hermes, Claude Code and Codex. On top of that, we also show our integrations with OpenClaw and Hermes through their existing hooks. A systematic breakdown can be found in Table 6 (Appendix D.1).

Weakness 2

“Lack of Independent Human Validation for Reference Labels: The ground-truth violation labels used to evaluate VAI on τ³-bench and ClawsBench lack rigorous independent validation… the absence of an independent human audit undermines confidence in the reference labels when interpreting both missed violations and false positives.”

We agree and have since removed that Dawid-Skene study and the five-labeler comparison. Three LLMs now label every run independently and we take their 2-of-3 majority as the label. We then sample 30 random runs per benchmark and have two human annotators independently label them, without seeing each other's labels or the LLM labels. Neither annotator had seen the traces before doing this annotation. Before resolution, the two annotators agree on 26/30 runs (Cohen’s kappa = 0.64) on Tau-Bench and 27/30 on ClawsBench (Cohen’s kappa = 0.52). The human annotators then discuss and resolve their disagreements and establish a human baseline. After discussion, the human annotators agreed on all annotations. We compare the LLM majority on those same 60 runs with the human baseline and they agree on 27/30 Tau-Bench runs (Cohen’s kappa = 0.71) and 24/30 ClawsBench runs (Cohen’s kappa = 0.14). The low Cohen’s kappa on ClawsBench is because violations are rare in ClawsBench and the human baseline labeled 5 violations and the LLM majority found only 1 of them. Section 5.1 and Appendix D.2 (Table 7) have been updated to reflect this.

Thank you again for your careful review and thoughtful feedback. We would like to follow up before today's discussion deadline to see whether our response has addressed your concerns. Please let us know if there is anything we can further clarify.

Dear Authors,

Thank you for your response and the effort invested in running instrumented evaluations on τ3\tau^3-bench and ClawsBench, as well as conducting an independent human annotation study.

To help me fully evaluate these new results, could you please clarify a few remaining details?

  1. False-Positive Rates and Denominators on the Updated Runs (Weakness 1 / Question 1): In Question 1, I asked for both the empirical detection rate and the false-positive rate (FPR) after integrating the augmented records. While you reported the 79.8%79.8\% detection rate, the corresponding FPR under the updated harnesses and newly registered promises is not mentioned. Furthermore, since adding an explicit approval tool in τ3\tau^3-bench alters the agent's action space and generates a new corpus of interactions, how did the total number of violating runs (the denominator) and non-violating runs change compared to the original uninstrumented runs? Please provide the full breakdown (Violations, Detected, and False Alarms, analogous to Table 1) for the newly executed runs, and clarify how AgentDojo is handled in the updated Table 1.

  2. Reliability of Ground-Truth Labels on ClawsBench (Weakness 2): While the agreement between the 3-LLM majority and the resolved human baseline on τ3\tau^3-bench is reasonable (27/3027/30, κ=0.71\kappa = 0.71), the agreement on ClawsBench is very low (24/3024/30, κ=0.14\kappa = 0.14), with the LLM majority missing 44 of the 55 human-identified violations and introducing 22 false positives. Given that the Standard60 cohort in ClawsBench has only 6060 runs in total—and your human annotators have already labeled 3030 of them—relying on the 3-LLM majority as ground truth for ClawsBench seems unreliable. How does VAI perform on ClawsBench when evaluated strictly against human-verified labels (or if human annotation is completed for the remaining 3030 runs)?

Thank you again for the clarifications.

Thank you for the follow-up. We answer both questions below.

1. False alarms and denominators on the updated runs

Table 1 in the revised PDF reports false alarms for both harnesses. The breakdown for the newly executed runs is:

BenchmarkModelViolationsDetected, original harnessDetected, updated harnessFalse alarms, originalFalse alarms, updated
Tau-BenchGLM-4.7-Flash12320 (16.3%)77 (62.6%)0/410/41
Tau-BenchQwen3-30B-A3B15129 (19.2%)87 (57.6%)1/121/12
ClawsBenchGLM-5.283 (37.5%)3 (37.5%)3/515/51
Newly executed runs28252 (18.4%)167 (59.2%)4/104 (3.8%)6/104 (5.8%)
All, with AgentDojo709451 (63.6%)566 (79.8%)20/298 (6.7%)22/298 (7.4%)

Both harness columns are computed on the same runs, so they share denominators. The original-harness column counts only detections from records the original harness logs.

Compared with the original submission, Tau-Bench went from 111 to 123 violating runs for GLM (53 to 41 non-violating) and from 154 to 151 for Qwen (10 to 12 non-violating). ClawsBench went from 4 to 8 violating runs (56 to 51 non-violating).

We did not run AgentDojo again and did not update its harness, so Table 1 reports it only under the original harness, on the same runs as the submission (266 and 161 successful attacks, with false alarms of 7/97 and 9/97, for GLM and Qwen). The "All" row counts AgentDojo with its original-harness numbers in both columns (Table 1, footnote a).

2. ClawsBench against human labels

As requested, here we score VAI against the human baseline, and against the LLM majority on the same 30 runs for comparison:

LabelsViolationsDetected, original harnessDetected, updated harnessFalse alarms, originalFalse alarms, updated
Human baseline52 (40.0%)2 (40.0%)1/251/25
LLM majority31 (33.3%)1 (33.3%)2/272/27

The two sets of labels give close results. VAI detects two violations under the human baseline and one under the LLM majority on the same 30 runs, with false alarms of 1/25 and 2/27. The ClawsBench result in Table 1 therefore holds under human labels.

ClawsBench contributes 8 of the 709 violations, so the headline numbers depend little on its labels. Without ClawsBench, VAI detects 448/701 (63.9%) with the original harnesses and 563/701 (80.3%) with the updated harnesses.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the paper and revisions

Thank you to the AC for your valuable time and consideration; we would like to provide a brief summary of the paper and discussion. In this work, we have proposed verifiable agent insurance (VAI), a framework for agentic accountability: in VAI, agent providers commit to agentic behavioral policies on a blockchain, along with a monetary escrow. In case of agent failures, users submit a claim, an independent verifier checks the violation against the committed traces, and the smart contract pays the user from the escrow account.

We thank all reviewers for their constructive feedback. We appreciate the reviewers acknowledging that our work addresses a practically important question (Reviewers TwNg, imYq), delivers a complete end-to-end implementation with harness integrations and Base mainnet tests (Reviewers TwNg, H5wj, imYq), and shows that the information needed to audit current agents is often embedded in natural language rather than in structured records (Reviewers TwNg, H5wj, imYq). Reviewer H5wj also valued our released policy compliance labels on two benchmarks. We have responded to each reviewer concern individually. Below, we summarize the most common comments, and how our revised PDF addresses them.

1. Our initial submission argued that by making agents output more structured objects to track user consent and actions, we could increase the detection rate of VAI. The reviewers pointed out that this projection was not verified experimentally.

We have run experiments to validate these projections experimentally. We updated the Tau-Bench and ClawsBench harnesses and ran both benchmarks again to measure the coverage of VAI if structured records were emitted. VAI detects 63.6% of violations with the original harnesses and 79.8% with the updated harnesses, which corroborates our earlier projection within a small margin (abstract, Section 1, Section 5.1, Table 1, Section 5.2, and Table 6 in Appendix D.1).

2. Our initial submission labeled runs with a committee of three LLMs, weighted by Dawid-Skene error rates estimated on a calibration set that two humans and the same three LLMs labeled. We validated this committee against the majority vote of those five labelers, in which the three LLMs already form a 3-of-5 majority, leading to a possible circular dependency.

We have since removed the Dawid-Skene study and the five-labeler comparison, in favor of an independent human evaluation of the LLM committee’s judgments. Three LLMs now label every run independently and we take their 2-of-3 majority. Two human annotators independently labeled 30 random runs per benchmark, and we report their agreement with each other and with the LLM majority (Section 5.1, Appendix D.2, Table 7).

3. There was a concern that a provider may tamper with traces prior to committing them to the blockchain.

To address this, we have introduced signed checkpoints. VAI now generates a key pair for each session and commits the public key on chain. The SDK signs a checkpoint before and after each tool call, after which the user’s device verifies and stores it. The user submits the latest signature when filing a challenge, so any alteration to the records by the provider is ruled against the provider. This adds only 49ms per tool call with OpenClaw and 62ms with Hermes. For comparison, our OpenClaw agent takes 3.8 seconds per tool call including model generation, so the overhead is about 1.3% (Section 3.2, Section 3.3, Appendix C.2, Algorithms 1–2, Appendix D.4, Table 8).

4. A few reviewers noticed that a user could submit a claim and have the escrow pool be drained by the time the violation is verified.

We now reserve payouts at filing time. When a challenge is filed, VAI sets aside the payout for that challenge, and if the reserve is insufficient, challenges are queued and funded in filing order (Section 3.1, Section 3.3, Algorithm 1, Appendix B.2).

Overall, we feel the reviewers’ comments were very valuable and have strengthened the work; we hope the new updates address these concerns. We are happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.