acceptodds
Under review as a conference paper at ICLR 2027

INFUSE: A Unified Surface Benchmark for Indirect Prompt Injection in Agentic Systems

Abstract

LLM-based agents increasingly execute long-horizon, end-to-end workflows involving shell commands, code modification, dependency installation, and interaction with remote services. This growing autonomy expands their exposure to indirect prompt injection (IPI), where adversarial instructions embedded in external content are interpreted as legitimate task directives. Existing evaluations, however, often focus on isolated malicious strings, single-step tool calls, or narrowly scoped agent configurations. We introduce INFUSE, an end-to-end benchmark for evaluating IPI throughout the lifecycle of long-horizon agent workflows. INFUSE comprises realistic tasks and a structured taxonomy of attack scenarios spanning five threat categories. Attacks are injected through realistic interaction surfaces encountered during task execution, extending evaluation beyond attacks placed only in initial prompts or isolated tool interactions. We evaluate 19 LLMs from the GPT, Claude, Gemini, Llama, Gemma, and Qwen families, producing 41,824 adversarial trajectories (26,220 without defenses) and 720 benign reference runs. Across the benchmark, mean attack success reaches 19.6% under LLM-based evaluation and 15.4% under heuristic verification, with per-model ASR-L ranging from 3.2% to 41.9%. Existing guard models provide limited protection, while prompt-level instructional hardening reduces mean attack success from 58.2% to 14.3%, revealing substantial but model-dependent security–utility trade-offs. Our code is available at https://anonymous.4open.science/r/infuse-agent-bench/README.md.

Primary area:
datasets and benchmarks
Keywords:
Indirect Prompt Injection, LLM Agents, Agent Security

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_qQ5TOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces INFUSE, a benchmark for evaluating indirect prompt injection in terminal agents during end-to-end task execution. It includes 40 benign tasks and 56 attack scenarios across multiple surfaces, including files, configuration files, and terminal output. The authors evaluate 19 models and collect 41,824 attack trajectories. Results are assessed using heuristic checks, LLM judges, sampled human annotations, and tests across agent frameworks. The main findings are that task completion can coexist with security failures, stronger model capabilities do not necessarily improve robustness, and defense effectiveness depends on the model and execution setup.

Critical strengths
  1. The evaluation reflects practical terminal-agent workflows. Attacks are embedded in normal operations, such as reading files, modifying configurations, and executing commands. This allows the benchmark to capture security failures that arise during multi-step interactions and is the paper’s strongest contribution.
  2. The paper demonstrates that task success can coexist with security failures. For example, GPT-5 achieves 94.3% task completion under attack while still having an ASR-L of 26.3%. This shows that an agent may complete the user’s task while also following malicious instructions, providing strong evidence for evaluating security separately from task completion.
Critical weaknesses
  1. Model robustness is not sufficiently separated from execution limits. For example, Claude-Sonnet-4.6 has an ASR-L of only 3.2%, but the authors note that its runs were truncated by the cost cap. A low attack success rate may therefore result from early termination rather than resistance to injection. In addition, only 22.4% of previously successful attacks remained successful when replayed in Claude Code and Codex CLI. These cases show that results depend strongly on execution conditions, making it difficult to infer model robustness directly from the main table.

  2. Defense gains are not sufficiently supported by verified attack outcomes. In the database migration task T19, for example, the LLM judge labels Gemini-2.5-Flash as compromised, while the heuristic verifier detects no target attack artifact: ASR-L is 1, but ASR-H is 0. Such behavioral deviation matters, but it differs in severity from achieving the attack objective. Although the paper reports both metrics, the main defense comparison focuses on ASR-L. It therefore remains unclear how much the reported reduction in attack success translates into a reduction in actual risk.

Questions for authors
  1. Does a low ASR reflect resistance to injection or an inability to continue execution? Using Claude-Sonnet-4.6 as an example, please report the frequency of early termination, including cost-cap exhaustion, and clarify whether its low ASR persists when sufficient execution budget is available.
  2. Do the defense conclusions hold when success requires achieving the attack objective? Please provide ASR-H or target-state verification results for the five main defenses. Does instructional hardening remain the most effective defense, and do cases flagged only by the LLM judge, such as T19, materially affect the conclusions?
Reviewer_4EWoOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces INFUSE, a benchmark for evaluating indirect prompt injection (IPI) attacks in long-horizon LLM-based agent workflows. Unlike existing benchmarks that focus on isolated tool interactions or limited attack surfaces, INFUSE evaluates attacks embedded in realistic execution environments, including files, configurations, terminal outputs, and archives.

The benchmark consists of 40 legitimate tasks and 56 attack scenarios across five security threat categories. The authors evaluate 19 LLMs from six model families, collecting 41,824 adversarial trajectories. They also investigate five defense strategies and study attack transferability across different agent harnesses.

The results demonstrate that current LLM agents remain vulnerable to indirect prompt injection, with an average attack success rate of 19.6% according to LLM-based evaluation. The authors further find that stronger models are not necessarily more robust and that existing defenses exhibit varying effectiveness and utility trade-offs.

Critical strengths
  1. Comprehensive benchmark design. The paper provides a structured evaluation framework covering multiple attack surfaces, threat categories, and realistic agent workflows. This broader coverage is a meaningful improvement over benchmarks focused primarily on isolated prompt injections.

  2. Extensive experimental evaluation. The evaluation covers 19 LLMs across different model families and includes a large number of adversarial trajectories. The experiments provide useful insights into the security vulnerabilities of current agentic systems.

  3. Evaluation of security and task utility. The paper considers both attack success and legitimate task completion, using heuristic verification and LLM-based evaluation. This provides a more comprehensive perspective on the trade-offs between security and utility.

  4. Practical analysis of defense strategies. The comparison of multiple defense approaches, together with the cross-harness experiments involving Claude Code and Codex CLI, provides useful insights into the effectiveness and limitations of existing security mechanisms.

Critical weaknesses
  1. Concerns about evaluation reliability. The main attack success metric relies on GPT-4o-mini as an LLM judge. Although the authors conduct human validation, the single judge achieves substantially lower agreement with human annotations than the multi-model committee (κ = 0.53 versus 0.92). This raises concerns about the reliability of the reported attack success rates and the resulting comparisons between models.
  2. Limited generalizability across agent architectures. Most experiments rely on Mini-SWE-Agent, which exposes a single Bash command interface. Although the authors include cross-harness evaluations using Claude Code and Codex CLI, these experiments cover only previously successful attacks on selected models and tasks. Consequently, it remains unclear whether the reported vulnerability patterns and model rankings generalize to more diverse agent architectures and real-world deployments.
Questions for authors

see weakness

Reviewer_wCe6OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes INFUSE, a benchmark for indirect prompt injection against terminal agents. It has 40 Dockerized software-engineering tasks with test-script verifiers and 56 hand-written attacks in five threat categories. An injection engine plants the attacks into files, configs, filenames, terminal output or archives. Agents run in Mini-SWE-Agent and are scored by a GPT-4o-mini judge plus heuristic checks. Across 19 models, mean ASR-L is 19.6% and refusals are rare. On a high-ASR subset, instructional hardening works best and guard models help little.

Critical strengths
  1. The setting is realistic. Agents act through real shell commands in Docker, and injections can arrive through runtime output like terminal logs and error messages, not just simulated tool returns. This is much closer to how Claude Code or Codex are actually used than AgentDojo or InjecAgent.

  2. The evaluation is large (19 models, 26k+ undefended runs), and the judge is checked against a cross-family committee and human labels. The trajectories are released, so the data is reusable even if the conclusions are modest.

Critical weaknesses
  1. Novelty is limited. The task format is basically Harbor/Terminal-Bench (task.toml, instruction.md, solve.sh, test.sh), and this is not cited. The agent is Mini-SWE-Agent, the benign/attack pairing follows AgentDojo, and all defenses are existing methods. Prior IPI work on coding agents, such as AIShellJack, which reported 41–84% attack success across 314 payloads, and QueryIPI, is not discussed. So the "first" claim is overstated, and the main findings mostly confirm what earlier benchmarks already showed.

  2. The paper presents surface-level attribution as its key advantage over prior work. However, the results never break ASR down by surface or compare static and dynamic surfaces. The main claimed contribution is therefore not actually demonstrated.

  3. ASR-L is hard to interpret. The headline numbers come from a single GPT-4o-mini judge that the authors' own human study finds unreliable (κ = 0.53). ASR is also not conditioned on whether the agent ever saw the payload; for example, Sonnet's 3.2% comes from the cost cap. Some attacks also look like normal repo instructions, such as pinning a version or adding a pip index. This makes the model rankings and the "capability does not predict robustness" claim shaky.

  4. The defense study is weak. It uses a subset picked for high ASR and only five mostly small models. It tests content-safety classifiers (Llama Guard, Qwen3Guard) rather than actual injection detectors like Prompt Guard. The Progent port also raises Sonnet's ASR-L from 6.5% to 27.1%, which points to either a faulty port or a metric that counts intent rather than effect.

Questions for authors
  1. [Clarification] Are the tasks original or adapted from Terminal-Bench/Harbor? How does INFUSE differ from AIShellJack and QueryIPI?

  2. [Clarification] Can you report ASR per injection surface (static vs. dynamic), and ASR conditioned on the agent actually reading the payload? This would affect my score the most.

  3. [Clarification] Do the Table 3 numbers and model rankings hold if re-scored with the three-judge committee?

  4. [Clarification] Why does Progent increase ASR-L for Sonnet-4.6, and does ASR-H move the same way? (Suggestion: compare against dedicated injection detectors.)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.