acceptodds
Under review as a conference paper at ICLR 2027

Margin-Optimized Narrative Attacks on Harness-Augmented Agents

Abstract

Language-model agents run inside harnesses that feed them retrieved documents and tool responses, some of which an attacker can edit. Guardrails against prompt injection check this text for injected instructions, unnatural wording, or forbidden actions, but a passage that has none of these can still change which candidate the agent selects at a later decision. Such passages are hard to defend against: their effect depends on the specific decision and context, so a defender would have to check every sentence from every source against every later decision. We introduce the Margin-Optimized Narrative Attack (MONA), which finds such passages by continuing an ordinary narrative and keeping the continuations that most shift the target decision. Across three models and five agent benchmarks, MONA steers decisions at about twice the rate of ordinary text. Guardrails built for explicit injection do not detect it, and guardrails that do suppress it also reject legitimate inputs or alter clean choices.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, agent harnesses, implicit attacks, adversarial robustness, small language models, tool use

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

This paper introduces Margin-Optimized Narrative Attack (MONA), a method for steering LLM-agent decisions through seemingly benign text rather than explicit malicious instructions. MONA uses a constrained beam search to generate fluent narrative continuations, selecting passages that maximize the log-probability margin of an attacker-chosen alternative over competing actions. The attack modifies a lower-trust input field without changing the agent's model weights or higher-priority instructions.

The authors evaluate MONA across three models, five agent benchmarks, and multiple harness configurations. MONA achieves approximately 19–22% targeted decision-promotion rates, compared with 5.8–11.1% for ordinary-text controls. Additional experiments investigate the relationship between harness performance and attack susceptibility, the effectiveness of existing guardrails, and the computational cost of attack optimization. The evaluation focuses on candidate preferences under recorded interaction histories rather than complete attacked trajectories.

Critical strengths

1. A useful formulation of adversarial influence beyond explicit instruction following.

The paper addresses an important distinction between prompt injection and contextual manipulation. An agent can make an attacker-preferred decision without encountering any explicit instruction to do so, which challenges defenses designed primarily around instruction-hierarchy violations. MONA makes this threat concrete through an optimization objective defined over candidate log-probability margins, while restricting overt instruction tokens and target-name cues.

The experiments provide evidence that this phenomenon is not confined to a particular domain or decision type. More importantly, the comparison with ordinary text demonstrates that the observed shifts are at least partly attributable to targeted optimization rather than arbitrary context perturbation.

2. Broad empirical coverage and attention to defense-induced behavioral changes.

The evaluation spans tool selection, environment actions, product selection, and database-table choices across different agent harnesses. This provides a useful view of how susceptibility varies with model and task configuration. The paper also goes beyond reporting attack success by examining whether guardrails perturb clean decisions. In particular, the input-marking experiments reveal that ostensibly protective transformations can independently change candidate rankings, an observation relevant to evaluating the utility of agent defenses.

Critical weaknesses

1. The primary success metric is an offline preference shift, not demonstrated agent compromise.

This is my main concern. MONA is optimized and evaluated against a fixed recorded interaction history, where only the injected text changes. The resulting ASR measures whether the target becomes the highest-scoring candidate, not whether an actual agent would select that action or subsequently deviate from its task.

Two additional aspects make the reported rates difficult to interpret as practical attack success. First, Appendix F restricts evaluation to decisions whose clean target margin lies within [-6, 0] nats, explicitly favoring alternatives already close to the model's preferred choice. Second, most benchmarks score only a shortlist of up to six candidates, even when the available action space is larger. A positive margin within that shortlist need not correspond to the preferred action under unrestricted generation.

Consequently, the reported 19–22% ASR cannot be interpreted as the probability of successfully redirecting an agent during deployment. The paper would be substantially stronger with attacked rollouts measuring actual action changes and downstream task outcomes, together with the proportion of all decisions satisfying the current eligibility criteria.

2. The baseline comparisons do not adequately isolate the benefit of the proposed optimization.

MONA uses hundreds of victim-model scoring queries per decision (approximately 390 under the default configuration), whereas the principal ordinary-text control averages only three randomly sampled sentences. This comparison establishes an improvement over unoptimized text, but not whether MONA discovers an unusually effective class of narratives relative to a comparable search budget.

A more informative baseline would perform best-of-N selection over fluent, length-matched narrative continuations using the same victim queries, but without MONA's beam-search objective. The comparisons with existing attack methods also have limitations: Appendix B explicitly notes that the optimized baselines are not compute-matched and that some implementations are adaptations rather than complete reproductions. Moreover, the injection judge is explicitly designed to detect instructions and unnatural text, making the near-complete rejection of overt-instruction baselines less surprising.

These limitations matter because the methodological contribution is not merely that optimization beats random text, but that the proposed construction offers a meaningful way to produce effective and difficult-to-detect contextual attacks.

3. The conclusions regarding guardrail effectiveness and over-defense are stronger than the evidence supports.

The paper argues that defenses capable of suppressing MONA also interfere with legitimate behavior, but much of this conclusion is based on proxy measurements rather than actual task degradation.

For example, the perplexity comparison contrasts fluent MONA narratives with original source excerpts that may contain code, identifiers, and structured data. Their different rejection rates therefore partly reflect a substantial distribution mismatch. Similarly, input marking can independently promote the attack target: for Qwen3-8B, Datamark produces 13.9% target promotion without an attack, compared with 17.9% MONA ASR under the same transformation. Reporting the latter as residual attack effectiveness without separating guardrail-induced flips obscures the attack's incremental contribution.

The task-relevance judge also rejects all default MONA passages, while evidence that it systematically harms legitimate task completion remains limited. Demonstrating that a particular defense rejects irrelevant text or changes a clean ranking is not equivalent to establishing an unavoidable security–utility tradeoff.

Questions for authors

1. Can the authors validate MONA through actual attacked rollouts rather than fixed-history candidate scoring?

In particular, how often does a positive target margin result in the agent actually selecting the attacker-preferred action, and how frequently does this alter final task success? Reporting attack success against the full candidate space and the fraction of decisions excluded by the clean-margin threshold would materially strengthen the interpretation of the results.

2. How does MONA compare with a compute-matched narrative-search baseline?

If an attacker samples and scores a comparable number of fluent continuations under the same lexical constraints, how much additional success does MONA's margin-guided beam search provide? This would help establish the algorithmic contribution independently of optimization budget.

3. Can the defense evaluation measure task-level utility rather than relying primarily on changes in candidate rankings or rejection rates?

A comparison of defended agents on complete benign and attacked trajectories would clarify whether the observed interference constitutes meaningful over-defense, and whether the claimed security–utility tradeoff persists under realistic execution.

6 comments
AuthorsOpenReview ↗

Rebuttal Part 1

Thank you for the detailed review. We address the interpretation of attack success, the baseline comparisons, and the guardrail evidence separately below.

W1-a / Q1. Does an attack found with a fixed history still work when earlier decisions are recomputed?

The fixed-history evaluation is designed to isolate the injected text's effect on a decision. We additionally rerun the pipeline on swe-bench: 60 of 75 attacks that succeeded under fixed-state scoring still selected the target when the pipeline was rerun from the beginning.

In the controlled evaluation, the task, surrounding interactions, candidate set, and target remain the same while the inserted passage changes. This lets us attribute the difference in candidate preference to the passage, without mixing it with differences elsewhere in the interaction.

The additional experiment uses Qwen3-8B on SWE-bench issues in a folder → file → function → line localization pipeline. We first run the clean pipeline and optimize a passage for a decision on its recorded path. We verify the resulting target selection against the full candidate set at that state. We then insert the same passage into the issue description and rerun the pipeline from the folder stage. Each stage makes a new selection, and that selection determines the candidates available at the next stage. We do not restore the clean path or force the target to remain available.

Attacked decisionSuccessful under fixed-state scoringStill selects the target after rerun
File2015
Function2623
Line2922
Total7560

The 60/75 result measures retention among fixed-state successes: 80% remain successful when earlier choices and subsequent candidate sets are recomputed. It measures localization decisions; unrestricted next-action generation is tested separately below.

W1-b. How many decisions satisfy the [−6, 0] clean-margin criterion, and does this criterion explain the advantage over ordinary text?

We counted eligible decisions in the underlying clean records, rather than only in the attack sample. The eligible fraction varies substantially across configurations, from approximately 50% in SWE-bench to more than 92% in WebShop/Act.

Benchmark / harnessModelDecisions in the window / valid decisionsFraction
SWE-bench / Agentless, tool decisionsQwen3-8B733 / 1,42751.4%
SWE-bench / Agentless, tool decisionsQwen3-14B557 / 1,11150.1%
WebShop / ActQwen3-8B84 / 9192.3%
WebShop / ActQwen3-14B89 / 9692.7%

The margin criterion selects decisions where an alternative is sufficiently competitive to study targeted influence. It is applied identically to MONA and the controls, which use the same decisions and targets. Being close to a decision boundary therefore does not, by itself, explain why MONA promotes the target more often than ordinary text.

W1-c / Q1. How often does a positive target margin become an actual target action under unrestricted generation?

We have directly tested this: Over 90% cases can get recovered.

We have completed a direct next-action generation test on ALFWorld/ReflAct states and WebShop/Act states using Qwen3-8B. We reuse the previously constructed passages and their original insertion locations, without reoptimizing them for this test. The model generates its next response greedily from the response boundary, with no forced action prefix or shortlist constraint during decoding. Unmodified replay reproduces the recorded action at 91% ALFWorld traj and 93% WebShop traj. They directly connect the scoring metric to generated behavior: on the tested cases, successful margin flips translated into targeted action changes even when generation was unrestricted.

W2-a. Does comparison with three ordinary-text controls establish more than sensitivity to arbitrary text?

The ordinary-text comparison tests whether MONA's passages promote the target more often than an ordinary insertion. On the full Qwen3-8B evaluation set, MONA achieves 21.5% target promotion, compared with 9.6% for ordinary text.

Both experiments use the same decision, insertion location, and target, with length-matched passages. We average the results of the ordinary-text controls. This directly tests the alternative explanation that inserting an arbitrary fluent passage would produce comparable targeted changes.

The result establishes an advantage over ordinary text. The compute-matched comparison addresses the separate question of how effectively the search uses its scoring budget.

AuthorsOpenReview ↗

Rebuttal Part 2

W2-b / Q2. Does margin-guided beam search outperform compute-matched best-of-N narrative sampling?

To isolate the contribution of margin-guided search, we compare MONA with best-of-N on the same 100 decision points. Both methods use the same generator, opening prefix, source context, lexical constraints, and maximum passage length. We match the actual scoring budget separately at each decision. Best-of-N independently samples complete continuations and selects the passage with the highest target margin. Configuration

ConfigurationMONA ASR[Counterfactual]BoN ASR[Counterfactual]Difference[Counterfactual]
Default narrative17.0%13.0%+4.0 pp

MONA outperform independent sampling in both configurations with the same scoring budget. The improvement would support the benefit of using decision feedback during narrative construction: promising partial continuations guide subsequent expansion, allowing the search to concentrate evaluations on passages that increasingly favor the target.

W2-c. Are adapted attack baselines comparable when their implementations and search budgets differ?

We would like to clarify that the adapted baselines are effective at promoting the target in our evaluation. They achieve 50.3–69.6% raw ASR, and their low after-judge ASR reflects the rejection of successful attacks by the judge. However, our method can still have a high ASR after the judge’s inspection.

The adaptations allow all five methods to operate on the same writable fields and attacker-selected decisions. Across the same 1,017 Qwen3-8B decisions, the results are:

AttackRaw ASRAfter-judge ASR
MONA21.5%21.3%
InjecAgent69.6%0.0%
AgentDojo65.7%0.0%
ToolHijacker58.1%1.3%
GCG50.3%0.0%

After-judge ASR counts attacks that both promote the target and pass screening. The comparison shows MONA’s advantage under this common defense: its successful passages largely remain accepted, while the successful baseline attacks are mostly rejected. Implementation details and method-specific budgets are documented in the appendix; the separate compute-matched comparison addresses search efficiency.

W2-d. Is the judge comparison informative if detecting explicit instructions and unnatural strings is expected?

We would like to clarify the main finding of this comparison: a judge that reliably detects the evaluated conventional attacks still misses almost all successful MONA passages. This is the advantage MONA is designed to demonstrate: targeted influence that survives screening even when recognizable attack instructions are blocked.

The same independent judge reduces baseline ASR to 0–1.3%, while MONA retains 21.3% from a raw ASR of 21.5%. Detecting explicit instructions therefore does not establish that a guardrail can detect ordinary-looking text that promotes an attacker-selected decision. MONA does not query this judge during search, so the result also demonstrates evasion without optimizing against the detector.

W3-a. Does the perplexity comparison merely reflect differences between fluent narratives and code or structured source text?

The original excerpts represent legitimate inputs that the evaluated harnesses must process, including code, identifiers, and structured tool outputs. Their different text distributions are therefore part of the practical defense problem: the tested perplexity thresholds reject this legitimate material more often than MONA.

At a threshold of 100, the filter rejects 38.5% of original excerpts but only 0.1% of MONA passages. Lowering the threshold to 20 increases attack rejection to 62.9%, while also rejecting 76.9% of original excerpts.

These results show that perplexity is a poor discriminator between MONA and legitimate content in the evaluated input channels. A stricter threshold catches more attacks, but also filters out more of the normal material supplied to the agent. The measured effect is rejection of legitimate inputs; downstream task utility is a separate evaluation.

AuthorsOpenReview ↗

Rebuttal Part 3

W3-b. Could residual success under Delimit or Datamark come from the guardrail itself changing the choice?

We would like to clarify that the guardrail-only changes are part of the overguard behavior we aim to measure. In our evaluation, input marking can alter clean choices while still allowing targeted attacks to succeed.

An effective defense should suppress malicious influence while preserving clean choices. To examine the latter, we apply each transformation to the original input without adding an attack passage. Target promotion in this condition comes from the guardrail itself. This interference with clean choices is the overguard behavior we report, and is a reason to evaluate these transformations beyond their effect on attack ASR.

The original aggregate ASR includes those guardrail-induced changes. We additionally isolate the attack's contribution by retaining only decisions with a generated passage where the guardrail alone does not promote the target. On these same decisions, we compare source-conditioned MONA with ordinary-text insertions:

ModelGuardrailEvaluated decisionsSource-conditioned MONAOrdinary-text control
Qwen3-8BDelimit9030.0%6.7%
Qwen3-8BDatamark9128.6%6.6%
Qwen3-14BDelimit9525.3%8.8%
Qwen3-14BDatamark8721.8%7.3%
Gemma-4-12BDelimit8011.2%5.8%
Gemma-4-12BDatamark8011.2%4.6%

These results concern the source-conditioned variant, evaluated separately from default MONA. Together, the two analyses identify both problems: the transformations can interfere with clean choices, and targeted attacks remain effective after the transformations' own effects are removed. Thus, these input-marking defenses do not reliably suppress the tested attacks while preserving clean decisions.

W3-c. If the task-relevance judge rejects every default passage, what remains of the defense claim?

The task-relevance experiment tests a defense designed to address ordinary-looking narrative attacks and whether MONA can adapt to it. It reveals two challenges: the screen can reject legitimate feedback, and source-conditioned MONA can still promote the target while passing the judge.

The judge receives the user task and an input document and checks for unrelated content. This catches the default MONA passages, whose narrative topic is unrelated to the task: all are rejected on the Qwen3-8B subset.

Applying this screen throughout a harness also requires evaluating the documents and tool feedback supplied during its interactions. We observed the judge rejecting an unmodified retail tool response reporting that a replacement item was unavailable. This legitimate feedback informs what the agent can do next, illustrating how relevance screening can overguard when its usefulness is not apparent from the task and one document alone.

We then test whether the attack can adapt its narrative to the input while retaining the target-margin objective. Context-based prefix MONA begins with a short prefix derived from the input, such as a room observation or product listing. Source-conditioned MONA additionally supplies the source document to the generator and uses a stronger fluency weight. Neither variant queries the judge during search.

Qwen3-8B configurationRejected by the task-relevance judge
Default MONA100%
Context-based prefix MONA81%
Source-conditioned MONA73%

Rejection falls to 81% and 73%, respectively. For source-conditioned MONA, 5% of the 100 evaluated decisions both promote the target and pass the judge, demonstrating successful evasion rather than acceptance alone. The experiment therefore identifies a defense that blocks the default attack, while showing both false positives on legitimate feedback and a path for adapting the attack without access to the judge.

Reviewer_qiKwOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Margin-Optimized Narrative Attack (MONA), a novel approach to manipulating LLM-based agents through seemingly benign natural-language passages without explicit malicious instructions or references to target actions. MONA employs a beam-search-based optimization strategy to generate fluent narratives that maximize the probability margin of attacker-selected decisions over competing alternatives. The authors evaluate MONA across three language models, five agent benchmarks, and multiple agent harnesses, demonstrating that it can effectively steer agent decisions while largely evading conventional prompt injection defenses. The paper further investigates the relationship between agent task performance and attack robustness, showing that improved task accuracy does not necessarily imply stronger resistance to such attacks. Finally, the authors evaluate several guardrail mechanisms and highlight the trade-off between attack detection and over-defense, where stronger filtering can interfere with legitimate agent behavior.

Critical strengths
  • Comprehensive Ablation Analysis: The paper provides a thorough ablation study investigating the effects of different search hyperparameters, narrative prefixes, and computational budgets on attack effectiveness. It also systematically evaluates various defense mechanisms, including input marking, perplexity filtering, and task-relevance checks, offering valuable insights into the robustness of MONA and the limitations of existing guardrails.
Critical weaknesses
  • Limited Novelty Compared to Existing LLM Jailbreaking Research: The core idea of optimizing seemingly benign natural-language content to steer model behavior has already been explored in prior research on LLM jailbreaking and adversarial prompt optimization. While this paper extends the idea to agent decision-making, the methodological novelty appears somewhat incremental. The authors could better clarify the fundamental differences from existing approaches and establish the unique contributions beyond applying similar techniques to agent settings.

  • Limited Model Coverage and Generalizability: The evaluation is restricted to three relatively small open-weight models (Qwen3-8B, Qwen3-14B, and Gemma-4-12B), without considering frontier closed-source models such as GPT, Claude, or Gemini. This limits the generalizability of the findings to real-world agent deployments. Moreover, MONA relies on access to token-level probabilities, raising additional concerns about its applicability to closed-source models with limited API access. Intuitively, frontier models' behavior should be very different compared with the tested models on such simple perturbations.

Questions for authors

N/A

4 comments

Thank you for the review. We will clarify the contribution and present additional evidence on larger, API-served models below.

W1. Is MONA an incremental application of jailbreaking or adversarial prompt optimization to agents?

We would like to clarify the main contribution of our work. MONA systematically constructs ordinary-looking narrative passages that promote a specific decision inside an agent harness, without explicitly requesting that decision or naming the target. To the best of our knowledge, this is the first study of narrative attacks on multiple agent harnesses.

Many prompt-injection attacks express the intended behavior through instructions, such as transferring funds, disclosing information, or selecting a particular tool. These requests give human readers and judge models a recognizable signal of the attack. Some token-level optimization methods instead produce conspicuous strings that provide another detection signal. MONA searches for fluent narrative text whose targeted effect is not stated in the passage itself. The generator continues a narrative, while feedback from the agent's decision determines which continuations are retained.

Our contribution is a systematic construction and evaluation of this attack surface across agent harnesses. It makes the gap between recognizing attack instructions and preventing targeted contextual influence experimentally testable.

W2-a. Do the findings extend beyond the three relatively small open-weight models?

We have completed additional evaluations on three API-served victims across ALFWorld, WebShop, and InterCode-SQL. ASR ranges from 13.8% to 25.9%, compared with 4.4–4.9% for ordinary-text controls.

VictimASROrdinary-text control
Qwen3-235B-A22B13.8%4.6%
GLM-5.318.3%4.4%
DeepSeek-V4.1-Flash25.9%4.9%

These are pooled rates over the three benchmarks. The API implementation samples constrained narrative continuations and selects them with the target-margin objective, using a 100-evaluation search cap per decision. It uses neither victim weights nor gradients. These experiments use recorded Qwen3-8B histories; we separately test the larger victim's own histories below.

The results extend the observed attack surface to a substantially larger model and additional model families through an API-only scoring interface.

W2-b. Could the larger-model results depend on using histories generated by a smaller model?

We additionally evaluated Qwen3-235B-A22B on its own ALFWorld trajectories. ASR is 32.1%, compared with 13.3% for ordinary-text controls.

The larger model generates the interactions used for this evaluation, and the attacks are constructed for decisions in those interactions. This addresses the concern that the effect might arise only when a larger victim is placed in contexts produced by a smaller model.

W2-c. Does access to token probabilities prevent MONA from working through model APIs?

Probability access does not inherently require access to model weights, and our completed API experiments demonstrate attacks through endpoints that support candidate scoring. The requirement is feedback on candidate probabilities, rather than weights or gradients.

Probability-returning interfaces also exist in commercial APIs. For example, OpenAI's Chat Completions API documents output-token logprobs and up to 20 alternatives through top_logprobs. These outputs differ from unrestricted candidate-sequence scoring, so compatibility depends on what the endpoint exposes. Our completed experiments establish feasibility on endpoints that provide the scores required by the current implementation. OpenAI API documentation

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer qiKw,

We would welcome your thoughts on our responses and the additional results shared above. If any concerns remain, we would be glad to discuss them further.

Thank you for your time and consideration!

Reviewer_qiKwOpenReview ↗

Response to Rebuttal

Thanks for providing the author's response and new results. This largely resolves my concerns. I have increased my score for it.

Dear Reviewer qiKw,

Thank you for considering our response and the additional results, and for updating your assessment. We’re glad we could address your main concerns. If there are any remaining points you’d like us to clarify, we’d be happy to do so.

Best regards, The Authors

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

The paper introduces MONA, an inference-time attack that inserts a benign-looking narrative passage into a writable input field of a harness-augmented agent. Using beam search, it generates fluent continuations that avoid overt instructions and candidate names, promoting an attacker-selected later decision. Across three models and five benchmarks, MONA steers decisions at about twice the rate of ordinary text, and the guardrails that suppress it also reject legitimate inputs.

Critical strengths
  1. The paper demonstrates a real and under-defended attack surface with concrete evidence: benign-looking narrative text can shift later agent decisions, and this effect survives several standard injection defenses.
  2. The evaluation is unusually thorough for an attack paper, spanning multiple models, benchmarks, and harnesses, and it systematically measures the cost each defense imposes on clean behavior, substantiating the claim that effective defenses are not selective.
Critical weaknesses
  1. Limited methodological novelty. The beam search maximizes the target margin, which follows the same optimization principle as the attack baselines the paper compares against. The genuinely new parts are only the lexical filters and the fluency term, so the technical contribution is thin beyond the observation that benign text can steer decisions.
  2. Narrow and overstated threat model. The success rate is measured on recorded, fixed histories, with pre-selected candidate shortlists and artificially near-tie decisions, and the authors admit the score measures only a local decision effect. However, the abstract still claims MONA steers decisions at about twice the rate of ordinary text, without mentioning this limitation.
  3. The experiments show evasion rather than stronger steering power. The adapted explicit-instruction baselines reach far higher attack success rates than MONA, and the comparison with ordinary text uses only random sentences. The baselines are not compute-matched, and MONA's success rate rises with a larger search budget, so the reported rates mix budget with attack quality.
Questions for authors
  1. The attack requires query access to the victim's token probabilities and a known candidate shortlist. How would MONA perform against a deployed agent with only black-box API access and an unknown candidate set, and would the reported success rates survive such a setting?
  2. The stronger context-based prefix variant raises the attack success rate substantially while partially evading the task-relevance judge. Has this variant been evaluated under the other guardrails, and does any defense remain effective against it?
5 comments
AuthorsOpenReview ↗

Rebuttal Part 1

Thank you for recognizing the attack surface and the breadth of the evaluation. We clarify the methodological contribution and provide additional evidence on execution and defenses below.

W1. Does MONA contribute more than a familiar search objective with lexical filters and a fluency term?

We would like to clarify the methodological contribution: MONA provides a systematic way to construct decision-targeted narratives without expressing the desired action or naming the target. Its construction separates the context that makes the passage read naturally from the agent context in which the passage is optimized to influence a decision.

A target-margin objective specifies which choice to promote, but does not by itself determine how the attack text should exert that influence. An unconstrained search can achieve the objective by directly requesting the action, naming the target, or producing conspicuous strings. These forms expose recognizable signals to an injection detector. MONA addresses the construction problem under constraints that remove those signals.

The generator continues a supplied narrative prefix in its own context. Each proposed continuation is then inserted into the agent's task context and scored for its effect on the target decision. Beam search retains partial passages using this decision feedback and their fluency in the generation context. The lexical constraints, fluency score, and separate contexts work together to construct text whose targeted purpose is not explicitly stated. The same formulation supports fixed-topic narratives, context-derived openings, and source-conditioned passages.

This contribution goes beyond observing that ordinary text can affect a model: it provides a method for finding such text for a specified decision and testing its effectiveness under defenses. In our completed comparison on 1,017 Qwen3-8B decisions, MONA retains 21.3% ASR after an independent injection judge, while InjecAgent, AgentDojo, ToolHijacker, and GCG retain 0–1.3%. The method therefore makes a concrete defense gap testable across agent harnesses: successful targeted influence can be carried by text without the conventional signals that the judge detects.

W2-a. Why evaluate fixed histories, and does the attack survive when earlier decisions are recomputed?

Holding the history fixed isolates the injected passage's effect by preventing other changes in the interaction from explaining the result. We additionally tested dynamic execution: 60 of 75 successful fixed-state attacks still selected the target when a Qwen3-8B we rerun from the beginning.

The controlled comparison uses the same task, recorded interactions, candidate set, insertion location, and target. Only the added passage changes. This makes the experiment directly informative about whether that passage can promote the targeted alternative.

To test whether success survives changes in the preceding interaction, we ran a separate SWE-bench experiment. The pipeline selects a folder, file, function, and line in sequence. We first optimize the passage at a decision on the clean path and verify success against the full candidate set. We then place the same passage in the issue description and restart from the folder stage. Every selection is recomputed, and each selected object determines the next stage's candidates. The clean path is not restored, and the target is not forced to remain available.

Attacked decisionSuccessful under fixed-state scoringStill selects the target after rerun
File2015
Function2623
Line2922
Total7560

W2-b. Does a positive margin within a shortlist translate into an actual generated action?

We have directly tested this: Over 90% cases can get recovered.

We have completed a direct next-action generation test on ALFWorld/ReflAct states and WebShop/Act states using Qwen3-8B. We reuse the previously constructed passages and their original insertion locations, without reoptimizing them for this test. The model generates its next response greedily from the response boundary, with no forced action prefix or shortlist constraint during decoding. Unmodified replay reproduces the recorded action at 91% ALFWorld traj and 93% WebShop traj. They directly connect the scoring metric to generated behavior: on the tested cases, successful margin flips translated into targeted action changes even when generation was unrestricted.

AuthorsOpenReview ↗

Rebuttal Part 2

W2-c. Are near-tie decisions artificially created, and does the abstract make the evaluation scope clear?

The margin criterion selects measured states from clean agent runs; we do not modify the task or model to create near ties. MONA and ordinary-text controls use exactly the same selected decisions and targets, so this selection does not by itself explain MONA's higher target-promotion rate.

The target is the clean runner-up, and the main evaluation selects decisions whose target margin lies in [−6, 0]. The proportion of valid decisions meeting this criterion varies substantially:

Benchmark / harness and decision typeQwen3-8B: eligible / validQwen3-14B: eligible / valid
SWE-bench / Agentless, tool733 / 1,427 (51.4%)557 / 1,111 (50.1%)
WebShop / Act, product84 / 91 (92.3%)89 / 96 (92.7%)

Valid decisions require the recorded choice to be represented in the candidate set, agreement between the clean scorer and that choice, and applicable token-boundary checks. These counts come from the underlying clean decision pools, rather than only the attacked sample.

W3-a. If explicit attacks have higher raw ASR, does MONA show only evasion rather than a meaningful attack advantage?

We would like to clarify that our comparison measures successful targeted manipulation that survives screening. MONA achieves 21.3% on this joint outcome, while the evaluated conventional attacks achieve 0–1.3% under the same judge.

The completed comparison uses the same 1,017 Qwen3-8B decisions, writable input fields, and targets across five agent benchmarks. An independent injection judge checks the passages for explicit attack instructions and unnatural attack text. MONA does not query this judge during search.

AttackRaw ASRAfter-judge ASR
MONA21.5%21.3%
InjecAgent69.6%0.0%
AgentDojo65.7%0.0%
ToolHijacker58.1%1.3%
GCG50.3%0.0%

After-judge ASR counts a success only when the passage both promotes the target and passes the judge, using all 1,017 decisions as the denominator. A passage that passes but does not promote the target counts as a failure. The result therefore establishes successful manipulation after screening, beyond passage acceptance alone.

The baselines' high raw ASRs show that their adapted implementations effectively promote the target in this setting. Their low after-judge ASRs arise because the judge rejects those successful attacks. MONA's advantage is that almost all of its successful passages remain accepted. Method-specific adaptations and search budgets are documented in the appendix.

W3-b. Does beating random ordinary text establish that MONA has a targeted effect?

The ordinary-text comparison tests whether an unoptimized insertion produces comparable target promotion. On the full Qwen3-8B evaluation set, MONA achieves 21.5%, compared with 9.6% for ordinary text at the same decisions and targets.

For each decision, we insert three approximately length-matched Wikipedia sentences at the same location used by MONA and average their target-promotion outcomes. This controls for adding ordinary text of comparable length to the same agent context. The higher rate for MONA shows that its constructed passages promote the specified target more often than ordinary insertions do.

This control establishes targeted influence beyond the effect of an unoptimized insertion.

W3-c. Does MONA outperform compute-matched narrative sampling, or do the results mainly reflect search budget?

To isolate the contribution of margin-guided search, we compare MONA with best-of-N on the same 100 decision points. Both methods use the same generator, opening prefix, source context, lexical constraints, and maximum passage length. We match the actual scoring budget separately at each decision. Best-of-N independently samples complete continuations and selects the passage with the highest target margin.

ConfigurationMONA ASR[Counterfactual]BoN ASR[Counterfactual]Difference[Counterfactual]
Default narrative17.0%13.0%+4.0 pp

MONA outperform independent sampling in both configurations with the same scoring budget. The improvement would support the benefit of using decision feedback during narrative construction: promising partial continuations guide subsequent expansion, allowing the search to concentrate evaluations on passages that increasingly favor the target.

AuthorsOpenReview ↗

Rebuttal Part 3

Q1. Can MONA work through a black-box API, and where does knowledge of the candidate set come from?

We have demonstrated attacks through scoring APIs on three additional victims, without access to their weights or gradients. Across ALFWorld, WebShop, and InterCode-SQL, the resulting target-promotion rates exceed ordinary-text controls for all three victims.

VictimTarget promotionOrdinary-text control
Qwen3-235B-A22B13.8%4.6%
GLM-5.318.3%4.4%
DeepSeek-V4.1-Flash25.9%4.9%

The API implementation samples narrative continuations and selects them using the target margin, with a 100-evaluation search cap. Candidate information comes from the environment's exposed actions, available tools, product listings, or database schemas. These results establish feasibility for API access with candidate scoring and visible candidates. An output-only interface with a fully hidden candidate set would require a different scoring procedure.

Q2. Does the stronger context-based attack remain effective under the other guardrails?

Yes. We evaluated source-conditioned MONA under Delimit and Datamark, and it retains an advantage over ordinary text under both transformations across all three models.

To isolate the attack's added effect, the comparison retains decisions with a generated passage where applying the guardrail alone does not already promote the target:

ModelGuardrailEvaluated decisionsSource-conditioned MONAOrdinary-text control
Qwen3-8BDelimit9030.0%6.7%
Qwen3-8BDatamark9128.6%6.6%
Qwen3-14BDelimit9525.3%8.8%
Qwen3-14BDatamark8721.8%7.3%
Gemma-4-12BDelimit8011.2%5.8%
Gemma-4-12BDatamark8011.2%4.6%

Source-conditioned MONA uses a context-derived opening, supplies the source document to the generator, and increases the fluency weight. It does not query the guardrails during search. The results show that the targeted effect persists after removing choices already changed by the input-marking transformation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Follow-up before the discussion closes

Dear Reviewers,

Thank you for the time you’ve put into reviewing our paper. As the discussion draws to a close, we’d welcome your thoughts on our responses and the additional experiments we reported. If any questions remain, we’d be happy to address them while the discussion is still open.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.