- Rating
- 2
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper introduces AdaLCPI, an indirect prompt injection attack that splits a malicious instruction into two incomplete fragments plus a generic reconstruction cue, embeds them in long filler content retrieved through an agent’s tools, and refines them with an OpenEvolve search driven by a graded LLM judge and natural-language execution feedback. Evaluated on seven models across Email, GitHub, and Slack, it reaches 61.4% macro-average attack success at 8k filler versus 32.8% for a matched Trojan Hippo-style baseline. Ablations indicate that fragmentation and long-context embedding are effective only when combined and under adaptive search.
Critical strengths
- I appreciate the experiment of this work, in this paper, the authors not only evaluate the results based on 7 models from 5 model families, but also consider hyperparameter sensitivity and ablation studies.
- I think the paper’s structure is easy to follow, especially for section 3.3, the authors not only clarifies the approach with algorithm demonstration, but also uses a case to show how the whole approach works.
- From Table 1, the performance of the proposed AdaLCPI is promising and decent, and can also be adopted to black box models.
- Regarding the approach, I like the idea of prompt reconstruction, also a graded judge returns a score plus natural-language feedback explaining why the attack failed, which guides the mutator far better than a single success/fail bit.
Critical weaknesses
- Despite the diverse model selection on both commercial models and open source models, the gap of these models are relative large. For commercial models (normally they will have stricter guardrails), only GPT families are tested while I am also curious on how the proposed approach performs on other models such as Claude and Gemimi. For open source models, testing on more frointer models such as GLM, DeepSeek and Kimi can also make this work more solid
- It would help to check whether the individual fragments remain harmless in isolation. This is currently enforced through prompting, but optimizing for attack success could gradually push the fragments toward nearly complete attack instructions. A classifier check or a small human review of successful candidates would give readers more confidence that the attack actually relies on reconstructing the fragments.
- I would appreciate some validation of the GPT-4o judge. Since its judgments both measure attack success and guide the optimization, the search might favor responses that satisfy the judge without achieving the intended attack. Reporting agreement with human labels on a subset would make the results more convincing.
- I like the idea of reconstructing instructions from separate fragments. The graded score and natural-language feedback also give the mutator more detailed information about partial progress and why an attempt failed, which is a sensible design choice.
- I have some concerns about the coverage of the evaluation. The main ablation in Section 4.2 focuses on the four models with the largest gains, leaving it unclear whether the same conclusions hold for models with smaller gains or different behavior. Moreover, the 81 task-objective pairs are constructed from only nine benign tasks, with three per environment, so the variety of underlying workflows remains limited. Extending the ablation to one or two of the remaining models and including more benign tasks would make the broader conclusions more convincing.
Questions for authors
- I understand the choice of two fragments as the simplest setting for reconstruction. Could you clarify why their insertion positions are fixed after a single random draw, and whether the results are sensitive to the number of fragments or their positions? A brief discussion or a small sensitivity analysis would be helpful.
- Following up on W1, how’s the performance of the approach on other models such as Claude, DeepSeek etc.? My main concern is the gap of the model (size) between the commercial models used and open source models used are large.
- GPT-5.6-Luna shows higher attack success rates in Email than in GitHub and Slack. Could you briefly discuss what drives this difference? Benign task completion rates without attacks and a few representative failure traces would help readers interpret the results.
- The ablation shows filler is essential, but does its content matter or only its length? Did you compare different filler types at the same length to separate length from diluting the malicious signal?
- In Figure 2(c), I noticed that the black patch with a robot icon covers the main content. I wonder if that is designed on purpose or it is a technical issue when the figure was exported?
5 comments
Response to Reviewer SJzk (Part II)
Fragment insertion positions and number of fragments (Q1): We use two fragments because this is the simplest setting that requires the agent to combine separate pieces of information. We sample insertion positions once and keep them fixed during each optimization run. This allows the search to focus on improving the fragments and reconstruction cue without simultaneously changing their locations. To study position sensitivity, we run additional experiments using previously successful AdaLCPI-8k payloads. We replayed these attacks at three newly sampled insertion positions without further optimization. The table reports the percentage of originally successful attacks that remained successful at each new placement. Pass@3 reports whether an attack succeeded at least once across the three placements.
| Model | Cases | Run 1 | Run 2 | Run 3 | Pass@3 |
|---|---|---|---|---|---|
| GPT-4.1 | 47 | 59.6% | 61.7% | 57.4% | 72.3% |
| GPT-5.1 | 53 | 54.7% | 50.9% | 58.5% | 79.2% |
| GPT-5.6-Luna | 7 | 28.6% | 42.9% | 28.6% | 57.1% |
| Qwen-3.6-27B | 58 | 34.5% | 32.8% | 39.7% | 58.6% |
| Ministral-3-14B | 50 | 14.0% | 10.0% | 10.0% | 22.0% |
| Gemma-4-31B | 74 | 58.1% | 54.1% | 58.1% | 83.8% |
| Muse-Glimmer-30B | 59 | 52.5% | 57.6% | 57.6% | 74.6% |
The results show that insertion positions affect attack success, but previously successful attacks can often succeed at different positions without further optimization. The sensitivity varies across models, with Ministral showing the largest reduction. We focus on two fragments in this work and leave the effect of using more fragments for future investigation.
GPT-5.6-Luna performance across environments (Q3): Table 18 reports that GPT-5.6-Luna achieves 25.9% ASR in Email, compared with 0% in GitHub and Slack under AdaLCPI-8k. This difference may arise from the different workflows, available tools, and actions required to achieve the attack objectives. For example, Email tasks involve reading and forwarding messages, while GitHub and Slack require different tool interactions. Our evaluation compares attack methods under the same model and environment configurations, rather than isolating which environment characteristics cause differences in attack success. Explaining this variation more precisely would require additional no-attack baselines and an analysis of individual failure traces. We therefore avoid attributing the difference to a single factor without further evidence. We will clarify this distinction in the revised version.
Effect of filler content and length (Q4): Our experiments focus on whether embedding fragmented instructions in long contexts improves attack success. To isolate this effect, we keep the filler content fixed during each optimization run while varying the attack representation. Table 3 shows that fragmented instructions without filler achieve 0% ASR, while adding 8k filler increases ASR to 75.3%. In contrast, complete instructions with the same 8k filler achieve only 25.3%. We further evaluate filler lengths of 2k, 4k, 8k, and 16k tokens across all seven models. These experiments show that long-context embedding is important for the effectiveness of fragmented attacks in our setting. We do not claim that the specific filler content has no effect. Comparing different filler types at a fixed length would help answer that separate question. However, such an experiment introduces another variable and requires additional optimization runs to account for differences between filler types. Our main finding does not depend on filler content being interchangeable, but on the observed interaction between fragmentation and long-context embedding. We will clarify the scope of this claim in the revised version.
Figure 2(c) (Q5): Hm, we don't see a black patch; that is strange. In any case we will update Figure 2.
Thank you
We would like to thank the reviewer again for their time and constructive feedback. We believe our responses and revised manuscript address the main concerns. We would be happy to clarify any remaining questions during the discussion period.
Response to Reviewer SJzk (Part I)
We would like to thank the reviewer for the constructive and detailed feedback. We also appreciate the positive assessment of our attack formulation and experimental results.
Evaluation on additional model families (W1, Q2): To address the concern about model diversity, we run additional experiments on DeepSeek-v4-pro, extending our evaluation beyond the original seven models. We evaluated AdaLCPI-8k against the same attack baselines. These results provide additional evidence that AdaLCPI's effectiveness is not limited to the original model families. We will include these results in the revised version.
Individual fragment validity (W2): We agree that it is important to verify whether optimized fragments remain incomplete rather than gradually becoming full malicious instructions. To investigate this, we used an additional classifier. The table reports the percentage of successful attacks where either fragment was flagged by the classifier. Across 348 successful attacks, only 1.1% had either fragment flagged by the classifier. This suggests that the individual fragments are rarely identified as prompt injections when examined separately.
We also extended the fragment-removal ablation in Table 4 to all seven models. The reconstruction cue alone achieves 0% ASR across all model-environment combinations. When only one fragment and the cue are retained, success rates range from 0% to 29.3%, substantially below the original attacks. While classifier results alone cannot establish semantic incompleteness, the removal experiments provide further evidence that combining information from both fragments is generally important for attack success. We will include these results in the revised version.
Validation of the GPT-4o Judge (W3): Our evaluation separates binary attack success from the graded feedback used during optimization. The binary Judge uses deterministic decoding and evaluates the agent's tool calls, their arguments, and the final environment state. An attack is counted as successful only when the specified attacker objective is achieved. Merely mentioning or acknowledging the injected instructions does not count as success. The graded Judge is only used when the binary Judge reports failure, so intermediate progress scores cannot be counted as successful attacks.
We agree that a human evaluation would provide an additional independent check. However, reliably labeling multi-step execution traces requires clear annotation criteria, multiple annotators, and adjudication to account for differences in human judgments. Such a study is time-intensive and was not part of our evaluation. We therefore use the same trace-based success criteria across all attack methods and will clarify this procedure in the revised version.
Broader evaluation and ablation coverage (W5): We agree that evaluating only the four models with the largest gains limits the scope of the main ablation. We therefore extended the fragment-removal experiments from Table 4 to all seven models, including GPT-4.1, GPT-5.6-Luna, and Ministral-3-14B. The results remain consistent in showing that the reconstruction cue alone is insufficient and that removing a fragment substantially reduces attack success in most cases.
To test whether our results extend beyond the original nine benign tasks, we also conducted additional experiments on AgentDojo. We evaluated 100 task-objective pairs drawn from four AgentDojo suites, with 25 pairs per suite, across all seven models. AdaLCPI-8k achieves 41.6% macro-average ASR, compared with 34.6% for Trojan Hippo-style, and obtains the highest ASR on five of the seven models.
These results provide further evidence that AdaLCPI can also succeed in an established benchmark with different agent workflows. We will include these additional experiments in the revised version.