- 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
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
| Configuration | MONA ASR[Counterfactual] | BoN ASR[Counterfactual] | Difference[Counterfactual] |
|---|---|---|---|
| Default narrative | 17.0% | 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:
| Attack | Raw ASR | After-judge ASR |
|---|---|---|
| MONA | 21.5% | 21.3% |
| InjecAgent | 69.6% | 0.0% |
| AgentDojo | 65.7% | 0.0% |
| ToolHijacker | 58.1% | 1.3% |
| GCG | 50.3% | 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.
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:
| Model | Guardrail | Evaluated decisions | Source-conditioned MONA | Ordinary-text control |
|---|---|---|---|---|
| Qwen3-8B | Delimit | 90 | 30.0% | 6.7% |
| Qwen3-8B | Datamark | 91 | 28.6% | 6.6% |
| Qwen3-14B | Delimit | 95 | 25.3% | 8.8% |
| Qwen3-14B | Datamark | 87 | 21.8% | 7.3% |
| Gemma-4-12B | Delimit | 80 | 11.2% | 5.8% |
| Gemma-4-12B | Datamark | 80 | 11.2% | 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 configuration | Rejected by the task-relevance judge |
|---|---|
| Default MONA | 100% |
| Context-based prefix MONA | 81% |
| Source-conditioned MONA | 73% |
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.
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.
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.
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.