acceptodds
Under review as a conference paper at ICLR 2027

Share-Borne AI Virus: Memory-Hopping Attacks Across LLM Agents

Abstract

Large language models are increasingly deployed as stateful assistants that retain information across interactions and use tools to read, modify, and create persistent artifacts. As these artifacts are shared between users, they form an indirect communication channel between otherwise independent assistants. We study a failure mode in which this channel enables self-propagating attacks. We introduce artifact-mediated propagation, where adversarial content introduced through an artifact (e.g. a report), is stored in an assistant’s persistent memory, reproduced in a subsequently created artifact, and acquired by another assistant that later reads it. We evaluate this process in temporal human–agent universes that model artifact exchange between independently operated assistants over time, measuring whether an attack survives successive hand-offs, how many hops it reaches, and how broadly it spreads. We find that attacks can propagate across multiple independent assistants and persist over extended interaction sequences. In larger simulated environments, even GPT-5.6 Luna exhibits substantial spread, reaching 60–80% of agents with propagation chains extending to eight hops. These results show that persistent artifacts can act as durable carriers of adversarial state, allowing attacks to outlive individual interactions and spread across isolated assistants.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
memory, prompt injection, adversarial robustness

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_cKKdOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

Prompt injection is a major security vulnerability in LLM agents, arising from the blurred boundary between untrusted data and instructions. While recent frontier models have become significantly more robust to conventional prompt injections, complete protection remains challenging due to the inherent ambiguity: documents may legitimately contain instructions, preferences, and workflow conventions that are difficult to distinguish from malicious directives.

This paper investigates a novel attack surface in which malicious instructions embedded in an artifact are incorporated into an agent's persistent memory and subsequently propagated through new artifacts consumed by other, independent agents. The authors employ iterative adversarial search to discover effective attack templates for individual models, effectively creating a cat-and-mouse game between attackers and increasingly robust models. Experiments across several LLMs demonstrate substantial rates of memory infection and multi-hop propagation, with considerable variation in robustness across models.

An important aspect of the experimental setup is endpoint-assisted propagation, which distinguishes two questions: (a) whether an attacker can induce an agent to retain malicious instructions and carry them into subsequent artifacts, and (b) whether agents can autonomously reproduce the complete malicious payload across multiple hops. The main experiments provide evidence for (a), while the reliable propagation of the full payload in (b) is facilitated by an attacker-controlled endpoint that regenerates the instructions at each hop. The paper additionally investigates propagation without this assistance and explores an independent classifier-based defense for detecting poisoned memories.

Critical strengths
  1. Novel and practically significant attack surface. The paper identifies an important vulnerability in multi-agent workflows, where an agent acting on behalf of one user creates artifacts that are subsequently consumed by another user's agent. This is an increasingly realistic mode of information sharing as autonomous agents become more prevalent. Since users are unlikely to thoroughly audit every AI-generated document or piece of code, malicious instructions can propagate across otherwise independent agents without attracting attention. The paper highlights a consequential security risk that is not adequately captured by conventional single-agent prompt-injection evaluations.

  2. Convincing empirical investigation and excellent presentation. The paper is exceptionally well written, with a clear threat model, intuitive examples, and a systematic experimental methodology. The authors develop an iterative search procedure for discovering effective attacks, evaluate the resulting templates on held-out scenarios across multiple LLMs, and characterize how malicious instructions persist and propagate over multiple hops. The substantial variation in vulnerability across models is particularly interesting. The additional experiments on endpoint-free propagation and memory-based detection further strengthen the investigation. Overall, the experimental evidence convincingly establishes the relevance of the proposed attack surface, even if some aspects of the propagation mechanism merit further investigation.

Critical weaknesses

None

Questions for authors
  1. Endpoint-free propagation. The main experiments rely on endpoint-assisted propagation, where an attacker-controlled service restores the malicious payload at each hop. Appendix D evaluates endpoint-free propagation only on DeepSeek-V4-Flash, with considerably weaker results. Have the authors evaluated endpoint-free propagation on the stronger models considered in the main experiments? How much does propagation degrade without this assistance?

  2. Infection versus execution of malicious behavior. The paper primarily measures whether malicious instructions persist in memory and propagate through artifacts. Appendix E extends this evaluation to security-critical goals, but does not appear to measure whether agents actually execute those malicious objectives. Have the authors observed examples of such downstream malicious behavior? I would expect the models' safety training to provide an additional layer of protection when a malicious instruction is translated into an actual action, even if it has already been stored in memory. It would be interesting to understand how often infection translates into successful execution of the adversarial goal.

3 comments
AuthorsOpenReview ↗

Thank you for the review [1/2]

We thank the reviewer for their strong positive assessment of the paper’s clarity, informativeness, and soundness. We appreciate their recognition of (1) the novelty and practical significance of artifact-mediated attack propagation and (2) the systematic evaluation of multi-hop propagation across models and held-out scenarios.

Summary of our response

We thank the reviewer for two questions that both led to new experiments. We now evaluate endpoint-free propagation on all four main models (Q1): three of the four propagate to a third assistant in 31–39% of runs, while GPT-5.6 Luna resists the attack. We also measure whether stored adversarial memories are acted on (Q2): once the poisoned memory is retrieved, agents carry out the attacker-intended action in 72–82% of goal-adjacent queries. Details follow below.


Q1. Endpoint-free propagation on the main models

Appendix D evaluates endpoint-free propagation only on DeepSeek-V4-Flash. Have the authors evaluated endpoint-free propagation on the stronger models considered in the main experiments? How much does propagation degrade without this assistance?

We thank the reviewer for this question. We have now evaluated endpoint-free propagation on all four main models, on the same 36 held-out test universes and with a single seed artifact per run.

An attack optimized for the prompt-only setting. To test an attack designed to propagate without any endpoint, we ran the same automated search for each model in the prompt-only setting, with the same budget of 250 evaluation jobs as an equal-budget endpoint-assisted search. We then froze each template and evaluated it on the 36 test universes.

Target modelEndpoint-assistedPrompt-only
S(1)S(1)S(2)S(2)S(3)S(3)S(1)S(1)S(2)S(2)S(3)S(3)
GPT-5.6 Luna0.250.190.080.200.000.00
Kimi-K2.60.830.750.640.570.420.36
GPT-OSS-120B0.670.500.420.680.380.31
DeepSeek-V4-Pro0.860.860.750.740.470.39

(Observed S(h)S(h) for both attack variants; templates from the same 250-job search. Endpoint-assisted: payload restoration, prompt-only: no endpoint. 36 held-out universes, two runs each.)

Endpoint-free propagation is stronger on three of the four main models than on DeepSeek-V4-Flash. With a prompt-only template, GPT-OSS-120B, Kimi-K2.6 and DeepSeek-V4-Pro reach a third assistant in 31–39% of runs, and chains reach hop 4 on all three. Propagation is weaker than with endpoint assistance, but the artifact–memory–artifact mechanism sustains several successive transmissions with no external service restoring the payload. GPT-5.6 Luna, the most robust model in the main experiments, resists the prompt-only attack: the seed infects a first assistant in 20% of runs, but no chain reaches a second hop.

We have added these results to Appendix E, which previously covered only DeepSeek-V4-Flash.

AuthorsOpenReview ↗

Thank you for the review [2/2]

Q2. Infection versus execution of malicious behavior

Appendix E shows that security-critical goals propagate, but not whether agents act on them. How often does infection translate into executing the adversarial goal, given the models' safety training?

We thank the reviewer for this question. Storing an instruction in memory is not the same as acting on it, so we have measured execution directly, in a new appendix (Downstream Impact of Injected Memories).

Setup. The security-critical goals we use in the paper are agent actions from Hidden in Memory (Pulipaka et al., 2026) [1]: operational sabotage, system hijacking, running attacker-chosen software, data exfiltration, and propagation to third parties. The paper also provides environments to test downstream impact, we use 100 environments from them. For each of the four main subject models, we place poisoned memory alongside benign memories and place the agent in the environment and evaluate. The queries given to the agent are benign user-queries which are slightly adjacent to the adversarial goal. The agent completes the task in a simulated OpenClaw environment, and an LLM judge reads the execution log to decide whether the agent performed the attacker-intended action.

Target modelAdversarial usage rate (%)
GPT-5.6 Luna76.0 ± 7.0
Kimi-K2.682.0 ± 7.5
GPT-OSS-120B72.0 ± 9.0
DeepSeek-V4-Pro80.0 ± 7.5

Rate at which the agent performs the attacker-intended action, given that the poisoned memory is retrieved into context. ± is the 95% bootstrap confidence-interval half-width.

Infection usually translates into execution. Once the poisoned memory is in context, every model carries out the attacker-intended action in 72–82% of queries. Like the reviewer, we expected safety training to act as a second barrier at this point, but in this setting the evaluated models usually believe the injected memory. We attribute this to how the instruction reaches the agent: it is retrieved as one of the user's own stored memories, next to benign ones, rather than arriving as an untrusted request, so the agent treats the action as a routine part of the task.

We describe this in Appendix G.

[1] Sidharth Pulipaka, Stanislau Hlebik, Leonidas Raghav, Sahar Abdelnabi, Vyas Raina, Ivaxi Sheth, and Mario Fritz. Hidden in Memory: Sleeper Memory Poisoning in LLM Agents. arXiv:2605.15338, 2026. https://arxiv.org/abs/2605.15338


We thank the reviewer again for these two questions, which prompted the endpoint-free evaluation and the downstream-impact study. We hope the new results address the concerns raised, and we are glad to provide further detail.

This is interesting. Thanks for the additional experiments!

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

This paper studies whether an attack can spread between separate LLM assistants through files their users share. An assistant first stores malicious instructions in memory, then reproduces them in a later artifact that another assistant reads. The main experiments use an attacker-controlled endpoint to restore the payload in outgoing files. The authors evaluate propagation on synthetic workflows with held-out goals and report infection rates and chain lengths across several models. Additional experiments examine propagation without the endpoint, larger networks, and memory screening.

Critical strengths
  1. The setting is relevant to assistants that remember information and create documents for others to use. Showing that an injection can survive across separate users and tasks is more informative than demonstrating a single compromised interaction.

  2. The distinction between remembering a goal and preserving the instructions needed to spread it is useful. Testing frozen attack templates on unseen goals and workflows also makes the results more convincing.

  3. The appendix includes results for security-critical goals and propagation without an endpoint, as well as attack-development costs. These experiments help characterize the conditions under which the attack works.

Critical weaknesses
  1. How much of the propagation depends on the endpoint restoring the attack? Appendix I.3 shows that the agent writes malicious content before contacting the endpoint, but does not show how well that draft would infect the next agent on its own. Appendix D tests prompt-only propagation with a different model from those in the main experiments. I would like to see a comparison on the same model and workflows, keeping the seed count and search budget fixed, with an unchanged-file endpoint, payload restoration, and no endpoint.

  2. The conclusion mixes results with different seed counts. In Section 6.2, the single-seed experiment goal-infects 38% of GPT-5.6 Luna assistants, while the larger-universe result of 60–80% uses three seeds. The conclusion nevertheless attributes the 60–100% range across models to a single poisoned artifact. Please correct this statement or provide the corresponding large-universe experiment with one seed.

  3. I am not convinced that the censoring correction is justified. Whether a run is censored depends on whether all agents are infected and how recently the final infection occurred. These conditions are related to the propagation process itself. This matters when 65 of 69 DeepSeek-V4-Pro runs are censored. The window-sensitivity experiment is helpful, but the analysis also needs to account for how many agents remain reachable and whether further exposure opportunities exist. The observed and extrapolated propagation rates should be presented separately.

  4. The practical conclusions would benefit from a more controlled evaluation. Search budgets differ substantially across models, making their final attack rates difficult to interpret as a comparison of robustness. The memory-screening results are offline classification tests. Could the authors report attack success against search budget and run the propagation experiment with screening or outbound restrictions enabled? The effect of these defenses on legitimate task completion is also relevant.

Questions for authors

see Critical Weaknesses

5 comments
AuthorsOpenReview ↗

Thank you for the review [1/5]

We thank the reviewer for highlighting (1) the practical relevance of attack propagation across users and tasks and (2) the distinction between retaining an attack goal and retaining its propagation instructions, alongside tests on unseen goals and workflows. We also appreciate their recognition of (3) the appendix results on security-critical goals, propagation without an endpoint, and attack development costs.

Summary of our response

We thank the reviewer for concrete suggestions, several of which we adopted directly. The main changes are summarized below and detailed in the point-by-point responses.

  • Endpoint dependence (Q1). On the four main models, we compared payload restoration with an endpoint that returns files unchanged, and ran a prompt-only attack found with a fixed 250-job search budget. Restoration matters most beyond the second hop. Without any endpoint, three of the four models still propagate to a third agent in 31–39% of runs.
  • Seed counts in the conclusion (Q2). We have revised the conclusion and the abstract: a single seed artifact goal-infects 35–99% of assistants, and the 60–100% range comes from the large universes with three seeds. In 10 of the 12 large-universe runs, a single seed's lineage accounts for more than half of all infections.
  • Censoring (Q3). Following the reviewer's suggestion, we replaced it with an estimator conditioned on exposure opportunities. The qualitative conclusions hold (per-hop continuation probability 0.76–0.90).
  • Controlled evaluation (Q4). We ran an equal-budget automated search for every model. We also tested two action classifiers on 200 outbound uploads from our runs: the inexpensive one is prompt-injected by the infected assistant and blocks only 29%, while Gemini-3.8-Flash blocks all of them. On 759 benign AgentDojo actions, the two classifiers block 0.26% and 0.13% of legitimate steps. With the inexpensive classifier enabled end to end, propagation still occurs (S(1)S(1)–S(3)S(3) of 0.89/0.69/0.53 on GPT-OSS-120B and 0.81/0.53/0.31 on Kimi-K2.6).

We hope these revisions address the reviewer's concerns, and we would be glad to clarify anything further.

Q1. How much of the propagation depends on the endpoint restoring the attack?

We thank the reviewer for this suggestion. We ran both comparisons on the four main models and the same 36 held-out test universes, with a single seed artifact in every run.

Unchanged-file endpoint versus payload restoration. To measure how far the assistants' own drafts carry the attack, we kept each model's frozen template from the main evaluation and replaced the attacker's endpoint with one that returns the submitted file unchanged. Everything else is identical: model, template, universes, seed artifact and harness. Since no new search is involved, the search budget is held fixed by construction. The next assistant therefore reads exactly the text the infected assistant wrote. This experiment isolates the effect of the endpoint.

Target modelEndpointFIF$^{goal}_a$FIF$^{full}_a$FIF$^{goal}_d$S(1)S(1)S(2)S(2)S(3)S(3)S(4)S(4)
GPT-5.6 Lunarestores payload (Table 1)0.350.350.330.710.550.410.27
returns file unchanged0.080.080.210.560.390.030.00
Kimi-K2.6restores payload (Table 1)0.560.550.570.810.730.680.50
returns file unchanged0.110.110.460.710.690.110.00
GPT-OSS-120Brestores payload (Table 1)0.810.800.810.950.880.830.60
returns file unchanged0.260.260.440.920.690.250.08
DeepSeek-V4-Prorestores payload (Table 1)0.990.950.971.000.940.940.79
returns file unchanged0.200.200.410.940.860.330.11

Restoration matters most beyond the second hop. First-hop survival barely changes (within the bootstrap intervals for every model), as expected, since the seed assistant is infected before any endpoint is called. The draft that the seed assistant writes right after reading the full attack is often enough to infect one more assistant: second-hop survival stays at 0.39–0.86, compared with 0.55–0.94 with restoration. Third-hop survival falls from 0.41–0.94 to 0.03–0.33, no chain reaches hop 4 on GPT-5.6 Luna or Kimi-K2.6, and the deepest chain on any model reaches hop 5. Goal infection falls to 8–26% of assistants, against 35–99% with restoration.

Current models are aligned against following injected instructions, and the template's exact wording, found through optimization, is what lets it propagate. Each time an assistant rewrites it in its own word, the attack loses part of that wording, and with it part of its ability to infect. The endpoint removes this loss by restoring the optimized text at every hop.

AuthorsOpenReview ↗

Thank you for the review [2/5]

Prompt-only attack: We have separately optimized an attack for each of the 4 models with no dependency on the end-point. We fix the budget of 250-evaluation jobs (we also separately optimized our endpoint-assisted variant to maintain this budget, we now have a fair comparison, more on this in response to Q4).

Target modelEndpoint-assistedPrompt-only
S(1)S(1)S(2)S(2)S(3)S(3)S(1)S(1)S(2)S(2)S(3)S(3)
GPT-5.6 Luna0.250.190.080.200.000.00
Kimi-K2.60.830.750.640.570.420.36
GPT-OSS-120B0.670.500.420.680.380.31
DeepSeek-V4-Pro0.860.860.750.740.470.39

Observed S(h)S(h) for both attack variants; templates from the same 250-job search. Endpoint-assisted: payload restoration, prompt-only: no endpoint. 36 held-out universes, two runs each.

With a prompt-only template, GPT-OSS-120B, Kimi-K2.6 and DeepSeek-V4-Pro all reach a third assistant in 31–39% of runs, and chains reach hop 4 on all three. Propagation is weaker than with endpoint assistance, but the artifact–memory–artifact mechanism sustains several successive transmissions with no external service restoring the payload. GPT-5.6 Luna, the most robust model in the main experiments, resists the prompt-only attack: the seed infects a first assistant in 20% of runs, but no chain reaches a second hop. These prompt-only templates come from the same fixed 250-job search, so their rates are a lower bound on what a determined attacker could reach.

We have added this discussion to Appendix E.

These results show that it is still possible for the virus to spread from agent-to-agent without the external dependency with S(3) being over 0.31 for 3 of the 4 models.


Q2. The conclusion mixes results with different seed counts

We thank the reviewer. The 60–100% range comes from the large universes, which use three seed artifacts, and we have revised the conclusion accordingly: "A single poisoned artifact can trigger multi-hop propagation, goal-infecting 35–99% of assistants across models in held-out workflows. In larger universes seeded with three artifacts, the attack reaches 60–100% of agents." The abstract now also states that the large-universe result uses three seed artifacts.

The two settings answer different questions. Table 1 measures how often a single artifact starts an outbreak. The large universes study how an outbreak develops once under way: its size, speed and depth. Because each large run is about ten times the size of a test universe, we ran each model once per universe, and used three seeds so that each run contains an outbreak to study; with one seed, many runs would end at the seed (4 of the 36 seed agents never convert). As stated in Appendix J (Appendix H in the submitted version), we treat these results as descriptive and do not compare them with Table 1.

The spread is mostly carried by one artifact. Attributing each infected agent to the seed whose lineage infected it (Appendix J) shows that the extra seeds act mainly as redundancy. In 10 of the 12 runs, a single lineage accounts for more than half of all infections. The largest single lineage alone reaches the following fraction of all 30 agents:

Target modelLargest single lineage, share of all agents (3 runs)All three seeds combined
GPT-5.6 Luna23–43%60–80%
Kimi-K2.667–80%77–97%
GPT-OSS-120B50–90%90–100%
DeepSeek-V4-Pro53–83%90–100%

These lineage shares describe the outbreaks we observed, not the outcome of a single-seed run.

AuthorsOpenReview ↗

Thank you for the review [3/5]

Q3. Is the censoring correction justified?

We thank the reviewer for this comment. We agree with the central point and have revised the analysis accordingly.

Before the specifics, one note on scope: none of the paper's headline results depend on the censoring correction. Table 1, the abstract and the introduction report only observed quantities (FIF and observed hop survival S(h)). The 8-hop claim comes from direct observation in the 30-agent universes. The censoring-corrected curve is a secondary analysis in Figure 3 and one paragraph of Section 4 only. Most of the paper rests on the data we observed empirically rather than a mathematical model.

Q3a. Censoring depends on the propagation process itself

In the submitted version, a run was censored if every agent was infected, or if the run ended within Δ=3 steps of the deepest hop. Both conditions depend on how fast the attack spreads. Censoring was therefore informative, which violates the independence assumption behind the Kaplan–Meier estimator. The reviewer is right that this matters most for DeepSeek-V4-Pro, where 65 of 69 runs were censored.

Revised estimator: condition on exposure opportunities, as the reviewer suggests. We no longer decide from elapsed time whether a chain stopped or was cut off. Instead we use what the logs record directly:

A chain at hop hh had a chance to reach hop h+1h{+}1 if an agent other than its author read an infected hop-hh artifact and afterwards performed a writing task. We estimate the per-hop continuation probability as

  • A chain whose final hop never had a chance contributes no failure. This includes chains where every agent was already infected, and chains whose artifacts were never read before the universe ended.
  • A chain that had a chance and did not advance counts as a failure, however much time remained.
  • No time window is involved, and whether a chain counts as stopped depends on whether an exposure opportunity existed, not on the outcome.

Results (108 test-universe runs per model):

Modelp^\hat p, time window (submitted)p^\hat p, exposure-based (revised)95% CIStopped chains with no chance at the final hopStopped chains that had a chance
DeepSeek-V4-Pro0.980.90[0.85, 0.93]6642
GPT-OSS-120B0.930.89[0.85, 0.92]6538
Kimi-K2.60.840.84[0.80, 0.87]4246
GPT-5.6 Luna0.710.76[0.70, 0.82]3146

(95% CIs: bootstrap over universes.)

As the reviewer anticipated, the time-window rule gave higher continuation estimates for DeepSeek-V4-Pro and GPT-OSS-120B: 42 stopped DeepSeek-V4-Pro chains had an exposure opportunity at their final hop and are now counted as failures. Under the revised estimator, our qualitative conclusions are unchanged. Once a clean agent is exposed, chains continue each hop with probability 0.76–0.90, and the surviving fraction halves every 2.5 (GPT-5.6 Luna) to about 6 (DeepSeek-V4-Pro, GPT-OSS-120B) hops. Adding the 12 large-universe runs changes p^\hat p by at most 0.02.

We report the per-hop breakdown (transitions that occurred / transitions with a chance) in a new appendix table. For DeepSeek-V4-Pro: 1→2: 102/103, 2→3: 102/102, 3→4: 85/99, 4→5: 48/72, 5→6: 15/18, 6→7: 6/6.

Q3b. Window sensitivity is not sufficient

Recomputing the submitted estimator for windows of Δ = 2–20 steps shows why: as the window grows, more runs are censored, and for Δ ≥ 20 every test run is censored, since test universes last at most 20 steps, so p^→1\hat p \to 1 by construction. The exposure-based estimator above replaces the window entirely.

Q3c. Present observed and extrapolated propagation separately

In the submitted version, Figure 3 already plots the observed S(h)S(h) (dotted), the corrected estimate (solid) and the geometric extrapolation (dashed) as separate curves. All tables report observed quantities only. The text could have kept this separation clearer: Section 4 reported a corrected hop-7 survival alongside the observed results. In the revision:

  • Observed (Table 1, now with 95% CIs): FIF, S(1)S(1)–S(4)S(4), and the deepest observed hops in the large universes (hop 7 reached in 3/3 DeepSeek-V4-Pro, 3/3 GPT-OSS-120B, 1/3 Kimi-K2.6 and 1/3 GPT-5.6 Luna runs).
  • Model-based (clearly labelled, Figure 3 and one paragraph): the exposure-based p^\hat p and the projection it implies.
  • We removed the sentence reporting corrected survival at hop 7. Figure 3's caption labels the dashed curves as extrapolations and points to the per-hop opportunity counts in the appendix.

No claim in the abstract or introduction relies on the model-based estimates.

Reviewer_fHzfOpenReview ↗
Rating
2(−1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies whether an attack planted in one shared file can spread between personal AI assistants that never communicate with each other. Each user has an assistant with private persistent memory, built on the OpenClaw harness with a MEMORY.md file. Users exchange ordinary files such as reports and notes, and each assistant acts only when its own user asks it to read or write a file. The attacker poisons one seed file and controls nothing else. The attack spreads if three steps keep happening. A reading assistant stores the attacker's goal and a propagation instruction in its memory. Later, during an unrelated writing task, it copies the goal and instruction into a new file. Another assistant reads that file and repeats the cycle.

Two attack versions are studied. In the main one, called endpoint-assisted propagation, an infected assistant sends each file it writes to an attacker-run server, which inserts a fresh full copy of the attack text into the file before it is saved. The assistant only needs to remember the goal and a short instruction to call the server. In the second version, prompt-only propagation, there is no server, and the assistant must reproduce the full attack from memory alone.

For evaluation, the authors build simulated settings called human-agent universes. There are 36 test universes across six domains and six workload levels, plus 12 development universes and 3 large universes with 30 agents each. For each of four target models, one goal-agnostic attack template is tuned on the development universes, frozen, and then tested on unseen goals and unseen universes. An LLM judge labels memories and files as infected or not. The paper reports the fraction of infected agents and files, hop-by-hop survival of attack chains, a correction for chains cut off when a simulation ends, and a breakdown of spread into an encounter rate, how often a clean read hits an infected file, and a conversion rate, how often such a read infects.

Main findings. The seed infects at least one assistant in 78 to 100 percent of runs. Spread differs strongly by model. DeepSeek-V4-Pro and GPT-OSS-120B infect most agents and many chains pass hop 4. GPT-5.6 Luna and Kimi-K2.6 spread less. In the 30-agent universes, the attack reaches 60 to 100 percent of agents and the longest chain reaches 8 hops. The prompt-only version, tested once on one smaller model, is much weaker, with second-hop survival of 0.30. On the defense side, a cheap classifier that screens memory writes flags all infected memories with few false alarms, and the paper discusses blocking outbound calls and screening widely shared files.

Critical strengths

2 major things that the paper stands out:

  1. First systematic evidence that adversarial state can hop between fully independent assistants through ordinary file sharing. Earlier worm demonstrations needed a channel the agents themselves control, such as direct messaging, shared memory, automatic email indexing, or a shared skill library. Here the only link between assistants is that users pass files around for their own reasons, each assistant keeps its own private memory, and the attacker controls one seed file and nothing after that. Showing that propagation still works in this setting, across four models, on held-out goals and held-out workflows, closes the last untested channel in this literature. This is consequential to my assessment because it changes what deployments must defend. If private memory plus everyday file sharing is enough, then shared documents become part of the security boundary, which is the paper's central message. The evidence behind the claim is also coherent at each step, from memory snapshots to upload logs to transmission trees, so it does not rest on one lucky trace.

  2. A measurement framework that turns worm anecdotes into quantities others can compare and defend against. The human-agent universes, the split between goal infection and full infection, hop survival with a correction for simulations that end early, and the breakdown of spread into encounter rate and conversion rate together form a testbed the community can reuse. It also produced findings that stay useful beyond this paper. Spread is limited by how often infected files get read, not by readers resisting. Infected files keep exposing new readers for dozens of steps, and a chain that looks dead can resume because its files stay in circulation. A few widely read agents cause most transmissions, so screening popular files is a plausible cheap defense. This is consequential to my assessment for two reasons. It is the part of the paper most likely to be used and cited, and the promised release of the universes, harness, judge prompts, and endpoint code lets others check and extend it. It is also what makes the paper's weaker parts repairable, since the framing and errata problems can be fixed without rebuilding the framework.

Critical weaknesses
  1. The central claim is stronger than the evidence behind it. The title and abstract describe a self-propagating virus, but every headline number comes from the endpoint-assisted variant, where an attacker-run server inserts a fresh full copy of the attack text into each file an infected assistant writes. The hardest part of self-replication, getting a working copy of the attack into each new file, is done by the attacker's infrastructure rather than by the assistants. The variant without the server is tested once, on one smaller model, with a template the search could not improve, and it reaches second-hop survival of 0.30 with chains dying out by about hop 4. The paper is honest about this in Section 4.1 and in the limitations, but the abstract and title never mention the server, and the phrase self-propagating attacks will be read as the unaided case. This is consequential to my assessment because it decides what the paper's claim actually is. As a study of server-assisted carry-forward, the evidence is solid. As a demonstration of a virus that spreads on its own between personal assistants, the evidence is thin. The paper as written claims the second while proving the first, and this cannot be fixed by better writing alone. It needs either prompt-only results on the main models or a visible narrowing of the claim.

  2. Every quantitative claim passes through a measurement pipeline with unresolved validity questions. The judge labels whether the agent wrote the payload during the current interaction, not whether the memory carries it, and Appendix J shows that 14 of 54 memories labeled clean still contained the full payload. It is not stated which definition feeds the main infection metrics or how the persistence analysis handles this. The judge belongs to the same model family as one of the targets, and it is itself the attack target in Appendix D, with the size of the manual audit unstated. The counts also do not fully reconcile, 65 of 69 censored runs against 72 expected runs, 30 held-out universes in Appendix D against 36 elsewhere, and a fourth-hop value cited to a table column that does not exist. There are only two runs per universe, Table 1 has no intervals, and the corrected survival estimates at deep hops rest on very few uncensored runs. This is consequential to my assessment because the paper is a measurement paper, so its value is its numbers. I believe the qualitative phenomenon, since the traces, upload logs, and transmission trees are coherent. But until the labeling definition, the audit protocol, and the run accounting are clarified, I cannot tell how much trust to put in the specific rates that carry the abstract, and a rebuttal answer here can move my score in either direction.

Questions for authors

Some of the questions that stand out after reading the paper:

  1. Judge labels. Which definition of infection feeds the main metrics, the judge's authorship label or actual content presence in memory? Appendix J shows that 14 of 54 memories labeled clean still carried the full payload. Given that, how was the persistence analysis in Appendix I.2 computed, and how many labels did the manual audit cover? Did you check for bias from using a DeepSeek-V4 family judge on a DeepSeek-V4-Pro target?

  2. Run accounting. Table 1 reports first-hop survival of 1.00 for DeepSeek-V4-Pro over 36 universes run twice, so 72 runs should enter the analysis, yet the text says 65 of 69 runs are censored. What happened to the other 3 runs? Related, why does Appendix D use 30 held-out universes rather than 36, and where is the fourth-hop value of 0.05 that the text attributes to Table 3?

  3. Table 5. The text says all seven metrics are equal or lower under security-critical goals, but the table shows full infection rising for two models, from 0.73 to 0.80 and from 0.71 to 0.89. Can you correct the sentence and explain the increase?

  4. Harness configuration. What were the exact exec and approval settings in the OpenClaw setup you evaluated? If approvals were off, can you justify that as a realistic default? The paper currently supports the outbound-access assumption by citing web fetch defaults, but web fetch is a read-only GET tool while the attack uploads files through exec, which has its own approval prompts.

  5. Scope of the claim. Is there any prompt-only propagation result on the four main models, even preliminary? If not, will you state in the abstract and title that the main results use an attacker-run server that restores the payload at every hop? This answer decides whether I read the paper as a virus demonstration or as a measurement study of server-assisted propagation.

  6. Censoring window. Using your existing logs, how do the survival estimates change if the censoring window covers the stall lengths you document in Appendix H, which exceed 20 steps, rather than 3 steps? And how many uncensored runs inform the hop 7 and hop 8 estimates?

  7. Search history. In what order were the models searched, and did the search agent's persistent research log carry over between models? This bears on whether the 37-fold cost difference measures model robustness or accumulated search experience.

  8. EVOMAL contrast. Which version and section of arXiv 2608.25776 contains the social-post extension that failed to reach a second hop? I could not find it in the public v1. If it is not public, please soften the claim.

11 comments
AuthorsOpenReview ↗

Thank you for the review [1/7]

We thank the reviewer for their positive assessment of the paper’s clarity and informativeness and for highlighting two strengths. These are (1) the first systematic evidence that attacks can spread between independent assistants through ordinary file sharing and (2) a reusable framework for measuring propagation and evaluating defenses.

Summary of our response

The two concerns the reviewer identified as decisive were the scope of the claim and the validity of the measurement pipeline. We address both with new experiments and revisions, summarized below and detailed in the point-by-point responses.

  • Scope of the claim (W1, Q5). We revised the abstract to state that the headline results use an attacker-run endpoint that restores the attack. We also ran two endpoint-free comparisons on all four main models. When the endpoint returns files unchanged, propagation degrades mainly after the second hop. With templates searched for the prompt-only setting, GPT-OSS-120B, Kimi-K2.6 and DeepSeek-V4-Pro reach a third assistant in 31–39% of runs, while GPT-5.6 Luna resists the attack.
  • Judge labels and audit (W2, Q1). FIF is computed from labels on each agent's final memory, and the persistence analysis from memory-content labels. A manual audit of 50 labels found no errors, and a judge from a different model family (Gemini-3.8-Flash) agreed on all 200 re-judged samples.
  • Run accounting and uncertainty (W2, Q2). We completed a third run of every test universe (108 runs per model), re-ran the runs that had stopped on infrastructure errors, and added 95% bootstrap confidence intervals to Table 1. Appendix E now uses the same 36 test universes as the main experiments.
  • Censoring (Q6). We replaced the time-window censoring rule with an exposure-based estimator that needs no window. The per-hop continuation probability is 0.76–0.90 across models, and claims about hops 7–8 rest only on direct observation in the large universes.
  • Table 5 (Q3). With full infection computed as defined in Section 3, the increase is not present, and we have revised the text.
  • Search history (Q7). We state the search order and explain that experience carried over between models. We also add an equal-budget automated search, in which GPT-5.6 Luna is clearly the most robust.
  • Harness and outbound access (Q4). We describe the permission settings and test two action classifiers on 200 endpoint uploads. An inexpensive classifier blocks only 29% because the infected assistant prompt-injects it, while Gemini-3.8-Flash blocks all of them.
  • EVOMAL (Q8). We have revised the description in the related work.

We hope these revisions address the reviewer's concerns, and we would be glad to clarify anything further.


A. Scope of the claim: W1 + Q5

Q5. Is there any prompt-only result on the four main models?

Scope of the claim: We emphasise that the endpoint-assisted attack is self-propagating because each successful transmission can induce a newly infected assistant to repeat the process during a later task, without further manual intervention by the attacker. The endpoint performs a specific part of that process: it restores the full attack prompt in the artifact the assistant submits. To make this dependence explicit, we revised the abstract to describe the endpoint alongside the GPT-5.6 Luna result. The reported 60–80% goal-infection rate and chains reaching up to eight hops come from this endpoint-assisted setting with three seed artifacts.

Extra experiments without endpoint-assistance. In the main results, infected assistants trigger each new hand-off during ordinary tasks, while the endpoint inserts a fresh copy of the full attack prompt (without requiring any attacker intervention during the process). Appendix K.3 shows that assistants write the adversarial goal and carry-forward instructions before calling the endpoint. Whether those drafts can infect the next assistant on their own is a separate empirical question which we address next…

In the existing paper, Appendix E established that prompt-only (no endpoint assistance) multi-hop propagation can occur on DeepSeek-V4-Flash. However, we have now run and included further experimental results on other models (and setups) to establish more generally the extent to which the prompt-only (no endpoint assistance) attacks can work: (see next comment block please)

AuthorsOpenReview ↗

Thank you for the review [2/7]

The same attack, without restoration.* We kept each model's frozen template from the main evaluation and replaced the attacker's endpoint with one that returns the submitted file unchanged. Model, template, universes, seed artifact and harness are identical to the main evaluation. The next assistant therefore reads exactly what the infected assistant wrote, which isolates how much the restoration contributes.

Target modelEndpointFIF$^{goal}_a$FIF$^{full}_a$FIF$^{goal}_d$S(1)S(1)S(2)S(2)S(3)S(3)S(4)S(4)
GPT-5.6 Lunarestores payload (Table 1)0.350.350.330.710.550.410.27
returns file unchanged0.080.080.210.560.390.030.00
Kimi-K2.6restores payload (Table 1)0.560.550.570.810.730.680.50
returns file unchanged0.110.110.460.710.690.110.00
GPT-OSS-120Brestores payload (Table 1)0.810.800.810.950.880.830.60
returns file unchanged0.260.260.440.920.690.250.08
DeepSeek-V4-Prorestores payload (Table 1)0.990.950.971.000.940.940.79
returns file unchanged0.200.200.410.940.860.330.11

(Payload restoration: three runs per universe, as in Table 1. Unchanged-file endpoint: one run per universe; one Kimi-K2.6 run stopped on an infrastructure error, leaving 35 universes.)

Without restoration, propagation degrades mainly after the second hop. The draft that the seed assistant writes right after reading the full attack often still infects one more assistant: second-hop survival is 0.39–0.86, against 0.55–0.94 with restoration. After that, third-hop survival falls from 0.41–0.94 to 0.03–0.33, and the deepest chain on any model reaches hop 5. Overall, goal infection falls from 35–99% to 8–26% of assistants.

Aligned models resist injected instructions, and it is the template's exact optimized wording that gets past this. Each time an assistant rewrites the attack in its own words, part of that wording is lost. The endpoint prevents this loss by restoring the full text at every hop.

Prompt-only attacks on the four main models. We ran the prompt-only attack, in which no endpoint is involved and each assistant must carry and reproduce the attack from memory alone, on all four main models. Each template was found with the same automated search and the same budget of 250 evaluation jobs as the equal-budget comparison (Q7), then frozen and evaluated on the 36 held-out test universes, two runs each.

Target modelS(1)S(1)S(2)S(2)S(3)S(3)
GPT-5.6 Luna0.200.000.00
Kimi-K2.60.570.420.36
GPT-OSS-120B0.680.380.31
DeepSeek-V4-Pro0.740.470.39

(Prompt-only attack, no endpoint. Templates from the same 250-job search. Full infection, observed S(h)S(h); 36 held-out universes, two runs each.)

With a prompt-only template, GPT-OSS-120B, Kimi-K2.6 and DeepSeek-V4-Pro all reach a third assistant in 31–39% of runs, and chains reach hop 4 on all three. Propagation is weaker than with endpoint assistance, but the artifact–memory–artifact mechanism sustains several successive transmissions with no external service restoring the payload. GPT-5.6 Luna, the most robust model in the main experiments, resists the prompt-only attack: the seed infects a first assistant in 20% of runs, but no chain reaches a second hop.

AuthorsOpenReview ↗

Thank you for the review [3/7]

B. Measurement validity: W2 + Q1, Q2, Q6

Q1. Judge labels, audit size, judge-family bias

We thank the reviewer for raising this. The judge-label wording in the earlier version could have been clearer. We clarify below:

How the judge works. The judge (DeepSeek-V4-Flash) is called once per agent per step. It receives the adversarial goal and attack template, the user's request, tool calls made (artifacts read/memories written), the agent's persistent memory after the step, and the agent's reply. It answers three yes/no questions: whether the memory contains the goal verbatim (Q1), the carry-forward instruction verbatim (Q2), or either of them paraphrased (Q3). At each step, the judge tracks if the agent has written the goal and carry-forward instruction at that particular turn. In addition, the judge labels a snapshot of each agent's memory at every step for whether it contains the payload.

At the end of each run, every agent's final memory is judged again (this time to only check if memory is infected) and agent-written artifacts are judged with the same questions on the artifact text.

Which label feeds the main metrics. FIF is computed from the judge's labels on each agent's final memory, that is, on whether the memory contains the payload, not on whether the agent wrote it in a particular step. In every run, the infected count reported in Table 1 equals the number of agents whose final memory is labelled as containing the goal.

Hop survival S(h)S(h) is built from infection events instead. The per-step labels identify the step at which an agent first saves the payload, and the judged artifacts identify which document carried it, so each infection is linked to the previous hop. Here the per-step label is the right one, because a chain needs to know when and from what an agent was infected, not only whether it ends up infected. Later steps labelled clean do not remove an agent from a chain, since a chain depends only on each agent's first infection.

Persistence analysis (Appendix K.2). This analysis uses the judge's labels on whether the memory still contains the payload. Per-step authorship labels would not make sense here: an agent saves the payload once, so those labels would drop to zero at every later step even though the payload is still in memory. Retention of about 1.0 is only possible because the labels describe memory content.

Manual audit. We manually audited 50 judge labels and found no errors. Labelling is an objective task: deciding whether a memory or document contains a given statement or instruction, verbatim or paraphrased. It is well within the capability of the judge model and leaves little room for model-specific bias. To check this directly, we re-judged 200 samples of DeepSeek-v4-Pro with Gemini-3.8-Flash, a judge from a different model family. The two judges agreed on all 200 samples.

Q2. Run accounting

The 3 missing DeepSeek-V4-Pro runs stopped on an infrastructure error: a file scheduled for transfer was not available on any reachable filesystem, so the harness stopped the universe. We re-ran these runs (see "Two runs per universe, no intervals in Table 1" above).

The fourth-hop value referred to a column that Table 3 does not contain, since Table 3 reports survival up to S(3)S(3). We have removed this reference. The 30 held-out universes in Appendix E was an error. Appendix E now uses the same 36 test universes as the main experiments, and we checked that its numbers match.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.