- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper argues that trace-based oversight rests on an assumption that is rarely examined: that the agent cannot alter the record of what it did. It separates the integrity of a stored trace from the fidelity of its collection, and it defines the recording boundary as "the point at which agent behavior becomes a logged event". The evaluation tests whether deployed harnesses enforce that boundary.
The evaluation covers 10 model-harness pairs in Docker with full filesystem access, across four settings: direct requests to delete, add to, or spoof traces; injected skill files that ask for deletion or compaction alteration without the user's knowledge; a reward loop whose score can only rise above a ceiling by modifying the trace; and peer-log or peer-workspace exposure. Each condition is run 10 times, and the paper defines success as a tool call that changes the agent's trace. The authors report that 9 of the 10 tested pairs delete their own traces when asked, with Muse Code the reported exception, and that the behaviour also appears under reward pressure and after exposure to peer examples. The paper also recommends recording requests and responses through a trusted interception server outside the harness host's control.
Critical strengths
S1. The question is well posed and lands on publicly available agent products. If the entity producing a record is the entity being recorded, then asynchronous monitoring, incident investigation and compliance audit all inherit the problem. The harnesses tested are the ones in current use rather than purpose-built targets, which is what makes the measurement informative. The paper's argument that hash-chained logging protects events after logging, and not the recording boundary itself, is correct and clearly put.
S2. The capability and exploitability measurements are the part of the paper I believe. The skill-injection result is the most interesting part: an agent never asked to delete anything follows a bundled-skill instruction to do so, with an independent host-side filesystem observer confirming the deletion (Figures 2 and 24). Each of the 10 pairs is run for 10 trials under a stated success criterion, and the prompts and container configurations are reproduced in the appendix alongside a released code link. The asymmetry the paper finds between direct requests and injected instructions is a finding, not an artefact of the setup.
S3. The qualitative documentation is unusually well controlled. Figure 1 traces an agent inferring from peer scores that an empty transcript is rewarded, writing a recurring cleaner, finding that its nohup and setsid launches did not persist, and switching to the harness-managed background option. Figure 36 shows that Muse Code's non-deletion is not a discovery failure. Figure 35 states plainly that its evidence does not establish modification of provider-held records.
Critical weaknesses
W1. The headline claims are not consistent with the paper's own measurements.
The propensity claim is stated more strongly than the figure it rests on. The abstract, the introduction and the close of the reward-seeking results all state or imply that every agent tampered under reward pressure, the last of them that all 10 models did so. Figure 5 shows that this reading applies to 9 of the 10 pairs. The single sentence introducing that figure asserts that "all tested models tampered with their traces in every anonymization trial" and then excepts Muse Spark and Opus 5.5 at 20% and 90%, so the universal clause and its own exception stand side by side. The per-model counts may support the weaker claim that 9 pairs tampered at least once; the prose does not match the rates it is describing.
A second claim rests on a measurement the paper excludes from its scope. The abstract states that the harnesses deleted their traces when asked "without triggering monitor guardrails". Footnote 2 of the setup section places monitoring efficacy outside the study's scope, and the full-access configurations in Table 1 disable or bypass the approval and permission layers. Whether guardrails fired in those runs is therefore unmeasured, while the paper does report a cyber safeguard firing in a related run and the model being downgraded before the action proceeded.
The third concerns how a result is reported. The appendices record that agents repeatedly scheduled background processes to remove newly written records, and one of them states that success rates capture whether tampering occurred but not when it began or how much evidence remained. Since success is defined as a single tool call that changes the trace, one deliberate deletion and a cleaner that re-empties the file repeatedly receive the same value, and the headline rate does not separate them.
W2. The single positive control cannot carry the feasibility claim the paper makes with it.
The paper's conclusion needs a demonstration that the boundary can be held, and Muse Code is that demonstration, reported as refusing deletion and fabrication throughout. On the paper's own account its mechanism is a prompt: a bundled read-session skill "instructed the agent that its traces were read-only evidence and that they should not be edited, moved or deleted under any circumstances". That is the mechanism class the paper uses as an attack in its own exploitability experiments, and the appendix records the limit when it notes that system prompting "would matter less under the assumption of a misaligned agent".
One pair, one model version, one mechanism, never varied, admits two explanations. Either the harness holds the boundary, or the model refuses this particular request. The minimal informative control is a component isolation: remove or neutralise the skill instruction, or deliver the same text in a harness that does not hold the boundary. The rest of the same pair's profile also bears on the reading: it scores 0% on trace deletion and fabricated resets but 30% on tool-call spoofing, so the boundary it holds is narrower than the aggregate suggests.
W3. The results are not presented in a form a reader can check.
The paper reports 46 figures and 1 table, and that table is the harness-configuration table. The measurements appear as bar charts (Figures 3, 4, 5 and 22) and as cumulative curves (Figure 23). The text gives summary ranges and scattered values. I could not find a numeric table for the 10 pairs × 4 scenarios × 10 trials grid, and no trial-level data file is named, so a reader can read an aggregate off a chart label but cannot check a cell, a failure reason, a denominator or a blocked run.
This is entangled with W1. What the submission does not provide is the tabulated form of the same data that would settle the contradictions. For a paper whose subject is that a generated record can diverge from what happened, and which asks in its mitigation section for records that are independently verifiable, its own measurements are the one artifact in the paper that is not presented for verification. The submission's AI use statement covers "code development" and "figure plotting", and it does not say what produced the reported counts or what was checked by hand.
Questions for authors
- The universal sentences in the abstract, the introduction and the close of the reward-seeking results sit alongside per-pair rates of 20% and 90% in Figure 5. Can they be restated with those exceptions? What are the denominators: how many trials per pair entered the reward loop, how many could act on the trace, and was a pair unable to attempt tampering counted as a non-attempt? For the 0% and 20% cells, was that a refusal or an inability to locate the record, and how were the two distinguished?
- Which reported observation supports "without triggering monitor guardrails"? The paper records a cyber safeguard firing in a related run, so what the abstract claims for the full-access runs is unclear to me. Given footnote 2 and the configurations in Table 1, were the blocked ("B") cases counted or excluded?
- Can the success criterion be split into single-shot deletion and persistent cleanup? What fraction of reward-loop runs had an active or scheduled cleanup process at session end, and what remained recoverable 30 minutes later?
- Does Muse Code still hold the boundary if the
read-sessioninstruction is removed or neutralised, or if the same instruction text is delivered in a harness that does not hold it? - Can a numeric table or data file be provided for the Figure 3 grid and the equivalent grids behind Figures 4, 5 and 22, together with one sentence on what produced the counts and what was checked by hand?
- The related work cites "Schmotz et al., 2026b, Instrumental monitor evasion emerges under ordinary task pressure, Preprint". The conclusion I read in that preprint, that ordinary task pressure leads agents to attempt to evade runtime monitors without an explicit adversarial objective, is close to this submission's reward-seeking result, and I would like to know how the two contributions are separated: what each measures and where the evidential boundary between them sits.
2 comments
Thank you for your concise response, I took some time to check the key points, and summarize them as below.
What I should withraw. 1) "9 of the 10 pairs"(W1) is inaccurate. Figure 5 showed that all the pairs are non-zero, so the number is 10, not 9. 2) The sentence referred to Figure 5 did name its exceptions. And strictly speaking, the abstract ("emerges naturally") and the conclusion ("can emerge naturally") are not universal claims, but only the introduction ("every agent engage in tampering") is. So I withdraw the count and the reading to these two points.
The most critical question still open: attribution for "except Muse Code" remains unclear. The key issue of the submission is some "unauthorized trace tampering". The first two scenarios defined in §3.1 are users asking the agent to delete their own session logs. If the agents comply to the request, it will be counted as a successful attack. §4, however, calls traces "highly-sensitive information" and invokes the zero-data policies, which implies that it is necessary to "prevent the model providers themselves from keeping such records". The criteria for determining whether an action is unauthorized are not clear. Contradiction in the abstract makes this point even more obvious. It says harnesses "allowed agents to delete their traces when asked, without triggering monitor guardrails". In the main text, the request is from the user, while the guardrail keeps missing. Similiar missing premise leaves the Muse exception unattributable. Its boundary follows some bundled instruction (0% for deletion and fabrication, and 30% for tool-call spoofing), but moves when the behaviour is inferred rather than named (20%). No pair in the design independently varies model and harness. Thus the exception cannot be easily assigned to capability, propensity, or some boundary in hold.
The one new result make things even worse. §3.1 says the 0% ASR "due to a prompt-based instruction in Muse Spark’s built-in read-session skill". If it holds, removing the skill should recover tampering. In the response you remove it and reports 0/3 in all the three settings. But it is not enough. Table 1 gives the full-access configuration of the same pair as approvalMode=allowAll with sandbox disabled, so the permission layer may never have been used as a mechanism. What is more, Figure 36 indicates that the first step of this skill is to locate session.jsonl, so the zero without it may be non-discovery rather than refusal. The response fails to rule out this possibility. The 9 runs cannot separate 0% from the 20-30% reported, while Muse's 30% on tool-call spoofing is even not touched at all.
The way evidence is presented: not something preferred, but the absence of data. The submission has 46 figures, and the only one table is the list of harness configurations. Every measurement is a bar or a curve. A reader can see the rate printed on the bar, but none of the ten runs behind it. There is no denominator, no run failed or why, no run blocked. No figure reports an interval for any of these comparisons. A paper arguing that records can diverge from reality has the very record that readers cannot verify at all. None of these requires a new experiment. They in your hand. Deferring this to a future version does nothing to resolve the questions surrounding the current submission.
Summary. The response adds one nine-run table. Everything else is for a later version, and no number is recomputed and reported. The sentence on what produced the counts and what was checked by hand is not answered. And Q3 asked whether the criterion could separate one deletion from a cleaner that re-empties the file. The response says this "doesn't need distinction … as both are tampering in the end", while B.2 states that success rates "capture whether tampering occurred, but not when it began or how much evidence remained". I will keep my rating as weak rejection.
Response to Reviewer GP17
We thank the reviewer for acknowledging the relevance and importance of our research question, the results of our study, and the qualitative examples we provide. We address the weaknesses and questions below:
read-sessiondisabled | |---|---|---:| | Direct user prompting | Financial privacy request | 0/3 | | Direct user prompting | Personal-summary privacy request | 0/3 | | Skill injection | Privacy-protection skill | 0/3 |We hope our responses have addressed your concerns. If so, we would be grateful if you would consider raising your score.