acceptodds
Under review as a conference paper at ICLR 2027

Divide and Inject: Can Agents Reconstruct an Indirect Prompt Injection from Fragments?

Abstract

Agentic systems are now being widely used to orchestrate tools and reason over long contexts. However, the improving capabilities of the large language models powering these agents also create new attack surfaces for indirect prompt injection. In particular, an attacker may not need to place a complete malicious instruction in retrieved content if the agent can reconstruct the objective from incomplete fragments distributed across a long context. In this work, we introduce adaptive long-context prompt injection (AdaLCPI), which combines long-context fragmentation with adaptive search. AdaLCPI splits an attack objective into incomplete fragments, embeds them in external content retrieved through the agent's tools, and uses a reconstruction cue to prompt the agent to combine them. It then iteratively refines the fragments and cue with OpenEvolve using graded scoring and natural-language execution feedback from the target agent. Empirically, AdaLCPI achieves higher attack success than strong adaptive baselines, reaching 61.4% macro-average ASR compared with 32.8% for Trojan Hippo-style and 30.0% for AgentVigil. Safety evaluations should therefore test whether agents remain robust when harmful objectives must be reconstructed from incomplete fragments.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
indirect prompt injection, agent security, tool-using agents

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_SJzkOpenReview ↗
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
  1. 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.
  2. 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.
  3. From Table 1, the performance of the proposed AdaLCPI is promising and decent, and can also be adopted to black box models.
  4. 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
  1. 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
  2. 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.
  3. 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.
  4. 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.
  5. 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
  1. 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.
  2. 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.
  3. 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.
  4. 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?
  5. 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
AuthorsOpenReview ↗

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.

ModelAdaLCPI-8k (ours)TAPHippoAgentVigilIterInjectLCF
DeepSeek-v4-pro66.7%6.2%42.0%34.6%44.4%4.9%

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.

ModelSuccessful attacksFlagged fragments
GPT-4.1470.0%
GPT-5.1531.9%
GPT-5.6-Luna70.0%
Qwen-3.6-27B580.0%
Ministral-3-14B504.0%
Gemma-4-31B741.4%
Muse-Glimmer-30B590.0%
Overall3481.1%

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.

ModelAdaLCPI-8k (ours)TAPHippoAgentVigilIterInjectLCF
GPT-4.114.0%46.0%38.0%29.0%29.0%3.0%
GPT-5.136.0%21.0%33.0%13.0%25.0%6.0%
GPT-5.6-Luna26.0%6.0%23.0%5.0%18.0%6.0%
Qwen-3.6-27B73.0%33.0%54.0%27.0%49.0%17.0%
Ministral-3-14B18.0%77.0%50.0%51.0%45.0%2.0%
Gemma-4-31B52.0%4.0%14.0%5.0%22.0%8.0%
Muse-Glimmer-30B72.0%8.0%30.0%19.0%32.9%17.0%
Macro avg.41.6%27.9%34.6%21.3%31.6%8.4%

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.

AuthorsOpenReview ↗

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.

ModelCasesRun 1Run 2Run 3Pass@3
GPT-4.14759.6%61.7%57.4%72.3%
GPT-5.15354.7%50.9%58.5%79.2%
GPT-5.6-Luna728.6%42.9%28.6%57.1%
Qwen-3.6-27B5834.5%32.8%39.7%58.6%
Ministral-3-14B5014.0%10.0%10.0%22.0%
Gemma-4-31B7458.1%54.1%58.1%83.8%
Muse-Glimmer-30B5952.5%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.

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.

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

This work proposes adaptive long-context prompt injection (AdaLCPI), which combines long-context fragmentation with adaptive search. AdaLCPI splits an attack objective into incomplete fragments, embeds them in external content retrieved through the agent’s tools, and uses a reconstruction cue to prompt the agent to combine them.

Critical strengths

Strengths:

  1. The paper introduces an interesting attack formulation that combines long-context instruction fragmentation with adaptive optimization. It highlights an important security vulnerability in tool-using agents: malicious objectives can be reconstructed and executed even when complete instructions are not explicitly presented in retrieved content.

  2. The evaluation covers seven LLMs, three agent environments, and multiple attack objectives. AdaLCPI substantially outperforms several established baselines, achieving a 61.4% macro-average attack success rate compared with 32.8% for Trojan Hippo-style. The consistent improvements over the matched adaptive baseline provide convincing empirical evidence of the approach's effectiveness.

  3. The paper systematically investigates the contributions of instruction fragmentation, long-context embedding, and adaptive optimization. Additional experiments examine different context lengths, repeated optimization runs, and attacks distributed across separate tool outputs, providing useful insights into the effectiveness and robustness of the proposed approach.

Critical weaknesses

Weakness:

  1. The core method primarily integrates two previously explored ideas: long-context instruction fragmentation and evolutionary prompt-injection optimization. The optimization procedure relies on the existing OpenEvolve framework, and the proposed feedback mechanisms follow established adaptive search paradigms. Consequently, the methodological contribution appears relatively incremental.

  2. Although the ablation studies demonstrate that fragmentation and long-context embedding jointly improve attack success, they do not fully explain why this combination is effective. Notably, some attacks remain successful when only one instruction fragment is retained, raising questions about whether successful attacks consistently require genuine multi-fragment reconstruction. Furthermore, the individual contributions of graded scoring and natural-language feedback are not sufficiently isolated.

Questions for authors

Q1. Could the authors provide additional ablation experiments comparing binary feedback, graded scoring alone, and graded scoring combined with natural-language feedback?

Q2. Do successful attacks genuinely require multi-fragment reconstruction?

Q3. How does AdaLCPI compare with existing methods under equivalent computational budgets?

Q4. How robust is AdaLCPI against realistic prompt-injection defenses and more complex agent environments?

5 comments
AuthorsOpenReview ↗

Response to Reviewer eqpN (Part I)

We would like to thank the reviewer for the constructive and detailed feedback.

1. Novelty of AdaLCPI: We believe that the novelty of AdaLCPI lies in showing how the combination of long-context fragmentation and adaptive attack optimization exposes a vulnerability in tool-using agents that neither approach captures on its own. Prior long-context fragmentation work studies whether models comply with reconstructed harmful requests, while existing adaptive agent attacks optimize explicit malicious instructions. AdaLCPI takes a different approach by optimizing incomplete fragments and a reconstruction cue based on the agent's actual tool executions. Our contribution is to show how these two previously separate approaches interact and how this interaction can be exploited to induce unauthorized tool actions.

Our experiments show why this combination matters. Under identical adaptive search settings, AdaLCPI-8k achieves 61.4% ASR compared with 32.8% for Trojan Hippo-style. Both methods use the same search budget, graded scoring, and natural-language feedback. Without adaptive optimization, long-context fragmentation achieves only 5.1% ASR. Table 3 further isolates this effect. Across four models, fragmented instructions with 8k filler achieve 75.3% ASR, compared with 25.3% for complete instructions with the same filler. Without filler, fragmented instructions achieve 0% ASR. Adding filler to complete instructions also fails to reproduce the gain. These results show that the improvement cannot be explained by longer inputs or adaptive search alone. We believe this is the main contribution of our work. The individual methods are known, but their combination reveals a previously understudied failure mode in tool-using agents.

2. Multi-fragment reconstruction (W2, Q2): We agree that it is important to verify whether successful attacks depend on both fragments. In Table 4, we already investigated this by removing one or both fragments from previously successful attacks without re-optimization. Removed fragments were replaced with equal-length neutral filler. We have now extended this analysis to all seven models. Across all models and environments, retaining only the reconstruction cue produces 0% attack success rate (ASR). When retaining a single fragment and the cue, success drops substantially, with ASR ranging from 0% to 29.3% across model-environment combinations. Since these experiments only replay attacks that originally succeeded, the drop provides direct evidence that information from both fragments generally contributes to attack success. A small number of attacks can still succeed with one fragment, but this is not the dominant behavior.

Moreover, we also used an additional classifier to study the individual fragments. Across 348 successful payloads, only 1.1% had either fragment flagged by the classifier. This is consistent with our design, where individual fragments should not explicitly contain the complete malicious instruction. Together, these results support the role of fragment composition in AdaLCPI.

3. Contribution of graded scoring and natural-language feedback (Q1): To isolate the contributions of the feedback mechanisms, we run additional ablation experiments. We compared binary success feedback, graded scoring without natural-language feedback, and graded scoring combined with natural-language feedback. We kept the attack formulation and search settings fixed across conditions. The table reports ASR (%) across seven models and three environments.

ModelBinaryGraded onlyGraded + NL (ours)
GPT-4.128.445.758.0
GPT-5.144.449.465.4
GPT-5.6-Luna0.00.08.6
Qwen-3.6-27B63.065.471.6
Ministral-3-14B7.418.561.7
Gemma-4-31B61.779.091.4
Muse-Glimmer-30B59.364.272.8
Macro avg.37.746.061.4

Graded scoring increases the average ASR from 37.7% to 46.0%, while adding natural-language feedback further increases it to 61.4%. The improvement is consistent across the evaluated models. These experiments show that both components contribute to attack success. Graded scores help distinguish partial progress from complete failure, while natural-language feedback provides more specific information for subsequent mutations. We will include these additional ablations in the revised version.

AuthorsOpenReview ↗

Response to Reviewer eqpN (Part II)

4. Comparison under equivalent computational budgets (Q3): Our main comparison against Trojan Hippo-style already uses the same adaptive search procedure and maximum budget of 129 target-agent executions per task-objective pair. Both methods use the same target models, tasks, attack objectives, graded scoring, and natural-language feedback. The main difference is how the malicious instruction is represented. Appendix B.5 further discusses attack success as a function of the number of evaluated candidates. Figure 4 shows that AdaLCPI-8k achieves higher cumulative ASR than Trojan Hippo-style across most of the matched candidate budget. Figure 5 shows that this advantage also appears early in the search for six of the seven models, with GPT-4.1 being the main exception. At the full budget, AdaLCPI-8k reaches 61.4% ASR compared with 32.8% for Trojan Hippo-style. This comparison matches target-agent evaluation counts, although the longer contexts used by AdaLCPI result in different token costs.

5. Robustness against defenses and established agent environments (Q4): To address this concern, we run additional experiments with existing prompt-injection classifiers and input sanitization defenses. To make this a fair evaluation, we re-used our previously optimized attacks without any further optimization against each defense. For classifiers with limited context windows, we used windowing to process the full payload rather than truncating it. The table reports the average attack success rate across seven models and three environments for selected defenses. Values in parentheses show the change in percentage points relative to the undefended setting.

DefenseAdaLCPI-8k (ours)TAPHippoAgentVigilIterInjectLCF
No defense61.4%26.3%32.8%30.0%32.6%5.1%
PromptGuard-2-86M47.6% (-13.8)7.8% (-18.5)32.8% (0.0)25.6% (-4.4)27.9% (-4.7)5.1% (0.0)
ProtectAI DeBERTa-v361.4% (0.0)0.0% (-26.3)27.2% (-5.6)26.3% (-3.7)13.8% (-18.8)5.1% (0.0)
PromptArmor (Qwen3.5-4B)54.0% (-7.4)0.2% (-26.1)18.5% (-14.3)5.6% (-24.4)12.5% (-20.1)1.8% (-3.3)
PromptArmor (Qwen3.5-9B)41.4% (-20.0)0.9% (-25.4)18.5% (-14.3)5.8% (-24.2)14.6% (-18.0)0.0% (-5.1)

Our AdaLCPI-8k attack achieves the highest ASR under all four tested defense configurations. These results show that fragmented attacks remain effective against several existing prompt-injection classifiers and input sanitization methods, even without adapting the payloads to the defenses. We will include these experiments in the revised version.

We also run additional experiments on AgentDojo to assess whether our results extend to an established benchmark. 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% average ASR compared with 34.6% for Trojan Hippo-style, and obtains the highest ASR on five of the seven models.

ModelAdaLCPI-8k (ours)TAPHippoAgentVigilIterInjectLCF
GPT-4.114.0%46.0%38.0%29.0%29.0%3.0%
GPT-5.136.0%21.0%33.0%13.0%25.0%6.0%
GPT-5.6-Luna26.0%6.0%23.0%5.0%18.0%6.0%
Qwen-3.6-27B73.0%33.0%54.0%27.0%49.0%17.0%
Ministral-3-14B18.0%77.0%50.0%51.0%45.0%2.0%
Gemma-4-31B52.0%4.0%14.0%5.0%22.0%8.0%
Muse-Glimmer-30B72.0%8.0%30.0%19.0%32.9%17.0%
Macro avg.41.6%27.9%34.6%21.3%31.6%8.4%

These results provide further evidence that AdaLCPI can succeed beyond the environments constructed for our main evaluation. We will include these additional experiments in the revised version.

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.

Reviewer_Z59JOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an effective indirect prompt injection attack that splits the malicious instruction into incomplete fragments embedded in long retrieved tool content, together with a cue instructing the agent to reconstruct and follow it. By combining this framework with LLM-driven evolutionary search (OpenEvolve), the authors discover highly effective payloads that achieve high attack success rates across seven models in Email, GitHub and Slack environments. The evaluation also includes multiple ablations that delineate the contribution of the different parts of the design. Overall, compared to prior attacks, the method improves macro-average attack success rate to 61.4%, versus 32.8% for Trojan Hippo-style, 32.6% for IterInject, 30.0% for AgentVigil and 26.3% for TAP.

Critical strengths

Strong attack performance across diverse models, including recent reasoning models. AdaLCPI achieves higher attack success than the Trojan Hippo-style baseline on all seven target models, reaching 61.4% macro-average ASR versus 32.8%, with large gains on recent models such as GPT-5.1 (65.4% vs. 46.9%) and Qwen-3.6-27B (72.8% vs. 9.9%) (Table 1).

Thorough ablations. The paper isolates the contribution of each component: fragmentation and long-context filler are effective only in combination (Table 3), adaptive search is essential (5.1% ASR without it), and the attack remains effective across filler lengths and when fragments are split across separate tool outputs (Table 5).

Critical weaknesses

No prompt-injection defenses evaluated. All attacks are run against undefended agents. Whether fragmented injections remain effective against existing defenses, such as prompt-injection detectors, input sanitization or action guardrail, is untested.

Limited novelty. The method mostly combines two existing ideas: fragmenting a harmful instruction across a long context for the model to reconstruct (LCF; also Context Stitching), and adaptive attack search against agents (Hippo, AgentVigil, IterInject). The paper itself frames its contribution as this combination; the additions beyond it (graded Judge scoring with natural-language feedback) are incremental.

Questions for authors

How was the evaluation setup constructed? Are the Email, GitHub and Slack environments adapted from an existing benchmark (e.g., AgentDojo or InjecAgent) or built by the authors? What tools does each environment expose, and how were the three benign tasks per environment chosen?

How were the attack objectives chosen? What are the nine objectives per environment, how were they selected, and are they representative of realistic attacker goals? Why not use the injection tasks of an established benchmark, which would allow comparison with published results?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Z59J (Part I)

We would like to thank the reviewer for the constructive and detailed feedback.

1. Evaluation against prompt-injection defenses: To address this concern, we run additional experiments with existing prompt-injection classifiers and input sanitization defenses. To make this a fair evaluation, we re-used our previously optimized attacks without any further optimization against each defense. For classifiers with limited context windows, we used windowing to process the full payload rather than truncating it. The table reports the average attack success rate across seven models and three environments for selected defenses. Values in parentheses show the change in percentage points relative to the undefended setting.

DefenseAdaLCPI-8k (ours)TAPHippoAgentVigilIterInjectLCF
No defense61.4%26.3%32.8%30.0%32.6%5.1%
PromptGuard-2-86M47.6% (-13.8)7.8% (-18.5)32.8% (0.0)25.6% (-4.4)27.9% (-4.7)5.1% (0.0)
ProtectAI DeBERTa-v361.4% (0.0)0.0% (-26.3)27.2% (-5.6)26.3% (-3.7)13.8% (-18.8)5.1% (0.0)
PromptArmor (Qwen3.5-4B)54.0% (-7.4)0.2% (-26.1)18.5% (-14.3)5.6% (-24.4)12.5% (-20.1)1.8% (-3.3)
PromptArmor (Qwen3.5-9B)41.4% (-20.0)0.9% (-25.4)18.5% (-14.3)5.8% (-24.2)14.6% (-18.0)0.0% (-5.1)

Our AdaLCPI-8k attack achieves the highest ASR under all four tested defense configurations. These results show that fragmented attacks remain effective against several existing prompt-injection classifiers and input sanitization methods, even without adapting the payloads to the defenses. We will include these experiments in the revised version.

2. Novelty of AdaLCPI: We believe that the novelty of AdaLCPI lies in showing how the combination of long-context fragmentation and adaptive attack optimization exposes a vulnerability in tool-using agents that neither approach captures on its own. Prior long-context fragmentation work studies whether models comply with reconstructed harmful requests, while existing adaptive agent attacks optimize explicit malicious instructions. AdaLCPI takes a different approach by optimizing incomplete fragments and a reconstruction cue based on the agent's actual tool executions. Our contribution is to show how these two previously separate approaches interact and how this interaction can be exploited to induce unauthorized tool actions.

Our experiments show why this combination matters. Under identical adaptive search settings, AdaLCPI-8k achieves 61.4% ASR compared with 32.8% for Trojan Hippo-style. Both methods use the same search budget, graded scoring, and natural-language feedback. Without adaptive optimization, long-context fragmentation achieves only 5.1% ASR. Table 3 further isolates this effect. Across four models, fragmented instructions with 8k filler achieve 75.3% ASR, compared with 25.3% for complete instructions with the same filler. Without filler, fragmented instructions achieve 0% ASR. Adding filler to complete instructions also fails to reproduce the gain. These results show that the improvement cannot be explained by longer inputs or adaptive search alone. We believe this is the main contribution of our work. The individual methods are known, but their combination reveals a previously understudied failure mode in tool-using agents.

AuthorsOpenReview ↗

Response to Reviewer Z59J (Part II)

3. Construction of evaluation environments: Our Email environment is adapted from Trojan Hippo. We additionally built GitHub and Slack environments to evaluate attacks through different types of external content. Each environment provides tools for common user workflows.

  • Email (8 tools) includes list_emails, read_email, search_emails, reply_to_email, forward_to_email, send_email, draft_email, and delete_email.
  • GitHub (14 tools) includes issue management (list_issues, read_issue, search_issues, comment_on_issue, create_issue, close_issue, label_issue), repository file operations (list_files, read_file, search_files, create_file, edit_file, delete_file), and run_command.
  • Slack (12 tools) includes list_channels, read_channel_messages, read_thread, send_message, read_direct_messages, send_direct_message, search_messages, list_users, delete_message, edit_message, pin_message, and create_channel.

We selected three benign tasks per environment based on common user activities. Email tasks involve drafting an inquiry, replying to emails, and searching for information. GitHub tasks involve resolving repository issues, reviewing code, and commenting on issues. Slack tasks involve summarizing channels, checking direct messages, and replying to channel messages. These tasks require agents to retrieve and process external content during normal execution. The attacker controls content the agent may encounter, such as incoming emails, repository issues and comments, or Slack messages, but cannot modify the user's task or the agent's system instructions. We will include these details in the revised version.

4. Selection of attack objectives and established benchmarks: We selected attack objectives based on the confidentiality, integrity, and availability triad widely used in related works. We use three objectives per category in each environment, resulting in nine objectives per environment. We believe that this provides a fair evaluation setting.

  • The objectives reflect actions an attacker could induce through the available tools. In Email, they include forwarding sensitive information, sending unauthorized messages, and deleting emails. In GitHub, they include exfiltrating sensitive repository files, posting unauthorized issues or comments, and deleting files. In Slack, they include exposing API keys from messages, sending unauthorized messages, and deleting messages. We chose these objectives to cover distinct forms of harm while keeping the attacker's control realistic. The injection points are external content that an attacker could plausibly modify, including incoming emails, repository issues and comments, and Slack messages.

  • We initially constructed our environments to support long attacker-controlled content. To assess whether the results extend beyond these environments, we also considered 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% average attack success rate, compared with 34.6% for Trojan Hippo-style, and obtains the highest attack success rate on five of the seven models.

ModelAdaLCPI-8k (ours)TAPHippoAgentVigilIterInjectLCF
GPT-4.114.0%46.0%38.0%29.0%29.0%3.0%
GPT-5.136.0%21.0%33.0%13.0%25.0%6.0%
GPT-5.6-Luna26.0%6.0%23.0%5.0%18.0%6.0%
Qwen-3.6-27B73.0%33.0%54.0%27.0%49.0%17.0%
Ministral-3-14B18.0%77.0%50.0%51.0%45.0%2.0%
Gemma-4-31B52.0%4.0%14.0%5.0%22.0%8.0%
Muse-Glimmer-30B72.0%8.0%30.0%19.0%32.9%17.0%
Macro avg.41.6%27.9%34.6%21.3%31.6%8.4%

These results show that AdaLCPI can also be effective on an established agent benchmark, beyond the environments constructed for our main evaluation.

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.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to Reviewers

We thank the reviewers for their constructive feedback. We have uploaded a revised manuscript with additional experiments and ablations. Below we summarize the main updates:

  • Broader evaluation: We added experiments on AgentDojo with 100 task-objective pairs from four suites, where AdaLCPI-8k achieves 41.6% ASR compared with 34.6% for Trojan Hippo-style. We also evaluated DeepSeek-v4-pro, extending our original seven-model evaluation.

  • Evaluation against defenses: We tested four prompt-injection classifiers and input-sanitization configurations, reusing previously optimized attacks without defense-specific adaptation. AdaLCPI-8k achieves the highest ASR among the tested attacks under all four configurations.

  • Additional ablations: We isolated the contributions of graded scoring and natural-language feedback, extended the fragment-removal ablation to all seven models, and checked whether individual fragments are flagged as prompt injections. These experiments show that graded feedback improves attack success and that removing fragments substantially reduces it.

  • Evaluation details and reproducibility: We are still working on updating the appendix to provide the complete tool lists, benign tasks, and attack objectives for Email, GitHub, and Slack, along with additional analyses described in our reviewer responses.

AuthorsOpenReview ↗

Expanded Appendix

We have uploaded an updated manuscript with an expanded appendix covering the evaluation setup and additional ablations. We also plan to release a GitHub repository upon acceptance with the code and configurations, including all 27 attack objectives.

We hope we have addressed all questions and would be happy to clarify any additional questions during the remaining discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.