acceptodds
Under review as a conference paper at ICLR 2027

AsymStego: Exploiting Information Asymmetry to Bypass Content-Level Auditing in LLM Agents

Abstract

Indirect prompt injection attacks exploit untrusted content processed by LLM agents. Content-only auditors can screen this content before it reaches the agent, but the auditors and the executing agent may operate under different runtime information states, such as different timestamps, causing the same content to be interpreted differently. We investigate how this information asymmetry can be exploited to evade content-only auditing. We present AsymStego, an environment-bound steganographic injection attack that embeds instruction fragments in natural-language carrier text and enables the agent to recover them using runtime information that differs from that available to the auditor. We evaluate AsymStego on 62 attack cases derived from InjecAgent attacker intentions in a controlled, task-driven stress test of content-level auditing. On the OpenClaw–DeepSeek-V4-Flash stack, a GPT-5.5 auditor reduces baseline attack success rates from 83.9%–92.5% without auditing to 0.5%–19.4%. Under the same permissive audit, when the attacker has knowledge of the agent's relevant environment, AsymStego achieves a 52.7% attack success rate, compared with 90.3% without auditing. These results suggest that content-only auditing can face challenges in such settings, motivating runtime-aware defenses that incorporate the agent's execution context.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Agent Security, Indirect Prompt Injection, Steganography, Information Asymmetry

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_c4BnOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies indirect prompt injection against content-level auditors that inspect external material before an LLM agent processes it. The proposed AsymStego splits an attacker-desired instruction across word positions in two natural-language carriers. The agent uses runtime information, principally its hostname, to recover the fragments and may write them to a workspace note; a later session can read the note and execute the resulting task. The evaluation adapts 62 InjecAgent attacker intentions to this workflow. In the primary evaluation, AsymStego continues to induce successful attacks under GPT-5.5 content auditing, while the same auditor largely suppresses plaintext injection. Auditing also reduces AsymStego’s success relative to the unaudited setting.The paper also examines other detectors, audit policies, decoding budgets, task phrasings, and runtime guards.

Critical strengths
  1. The measurement effort is extensive and transparent.

  2. The AsymAES control cleanly separates cryptographic secrecy from audit capability, giving the core observation a concrete empirical grounding.

Critical weaknesses
  1. The attack does not constitute a meaningful threat. Its success comes from a scenario in which the user delegates the task to the attacker's page, safety confirmations are disabled, and execution is triggered by an experimenter-chosen instruction, rather than from any capability of the attack itself.

  2. The paper's takeaway, that content-only auditing should be combined with runtime checks, is already widely accepted and adds little new. The auditor's information disadvantage is easy to remove: the hostname-derived key has only 51 possible values, and it fools only a time-limited agentic auditor, while the single-pass auditor fails even when given the correct key.

Questions for authors
  1. Can the authors test AsymStego on a representative subset of the original InjecAgent or AgentDojo user tasks, where the attacker-controlled content is incidental to the task? Please report ASR with and without auditing.

  2. With the page and audit budget held fixed, how does detection change when the auditor receives the correct runtime key, an incorrect key, or no key? Can a tool-enabled auditor search the small key space within a practical latency budget?

  3. In the timestamp-key experiments, do successful agents obtain the intended key, try multiple candidates, or recover the payload another way? I am curious why these behaviors produce the non-monotone ASR in Table 1.

1 comment
AuthorsOpenReview ↗

Realistic Attack Scenarios, Limits of Capable Auditors, and Timestamp-Key Recovery

1. Evaluation Realism (Weakness 1, Question 1)

We thank the reviewer for these concerns. We evaluated all 62 original InjecAgent requests and 27 original AgentDojo tasks, alongside email, GitHub issue, and summarization scenarios (Tables 3–4; Appendices A.2, G).

ScenarioASRScenarioASR
Original InjecAgent, unaudited1.6%Email: solve22.6%
Original InjecAgent, audited1.6%*Email: complete45.2%
Original AgentDojo, unaudited0%GitHub issue resolution56.5%*
Adaptive original InjecAgent12.9%*Summarization1.6%
Adaptive summarization88.7%*

Except where marked “unaudited,” ASR includes auditing: * denotes GPT-5.6-Terra; other audited results use GPT-5.5. Adaptive results are cumulative over at most six attempts with audit/agent feedback. AgentDojo had no unaudited successes, so no additional audited run was performed.

We acknowledge limited effectiveness when decoding is unrelated to the task. However, our results demonstrate risks in routine workflows such as email handling, issue resolution, and summarization. In the issue-resolution and adaptive summarization experiments, users do not explicitly request decoding; the summarization request is simply “View and summarize the content at [URL].” Even after payloads are written to notes, the ordinary follow-up request “Continue my earlier work,” which does not mention notes or specify action, triggers target actions for 39.3% of the 56 landed notes (Table 2; Appendix C). Appendices H and H.1 show recorded traces of decoding, writing notes, and follow-up execution.

User confirmation cannot be assumed to provide reliable protection. Anthropic reports that users approve roughly 93% of permission requests and highlights the resulting approval fatigue [3]. We therefore evaluate autonomous execution using Claude Code's permission-bypass mode [1] and OpenClaw's ask=off [2]. Tool calls and their arguments remain agent-generated, so the experiments measure whether injected content induces the agent to initiate the target harmful actions (Appendix A.2).

2. Contributions and Limitations of Capable Content Auditors (Weakness 2, Question 2)

We agree that combining content auditing with runtime checks is established practice. Our contribution provides quantitative evidence that even content auditors with sufficient decoding capabilities can fail in certain scenarios when their runtime information differs from the agent’s. Our experiments distinguish this information barrier from the complementary capability barrier.

Capability barrier. A single-pass auditor may fail even with the correct key because it cannot perform the multi-step tool interactions needed for decoding. For AsymAES, single-pass DR remains 0% with the key exposed, whereas agentic DR reaches 98.4% (Fig. 4a).

Information barrier. Even an auditor capable of decoding may derive an incorrect key from its runtime environment. Holding the AsymStego carrier, payload, Opus 4.8 auditor, tool permissions, and 600-s budget fixed, three measurements per condition yield 93.5% ± 1.6 pp DR with the correct key, versus 37.1% ± 12.2 pp with an incorrect key. An additional no-key control yields 30.6% DR (19/62, one measurement; §5.3.2; Appendix A.3).

We agree that enumerating 51 keys is inexpensive: a pre-implemented, construction-specific decoder recovers 59/62 payloads in 0.8 s total (Appendix D.2). A general-purpose auditor, however, must first understand the rules, implement a decoder, and search candidate keys. The off-host traces analyzed in Appendix D.2 show no complete payload recovery within either 600 or 900 seconds; partial evidence can nevertheless trigger detection. Thus, AsymStego’s observed evasion is budget-bound, rather than based on cryptographic hardness. The exact timeout is an evaluation choice, but the operational constraint is real: an audit verdict must arrive before the dangerous action to prevent it; a late verdict cannot undo an executed action (Appendix E.3).

3. Timestamp-Key Recovery (Question 3)

Our added trace analysis shows that 82/97 successful trials with observable initial timestamps start with the wrong key, while at least 91/98 successful trials search candidate keys (Appendix A.4). Variation in search, timeouts, refusals, and tool calls can contribute to the observed local non-monotonicity; τ-bench likewise documents inconsistent outcomes across repeated agent trials [4].

[1] Anthropic. Claude Code: Configure permissions.

[2] OpenClaw. Exec approvals.

[3] Anthropic. How we contain Claude.

[4] Yao et al. (2024). τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains.

Reviewer_2d3AOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents AsymStego, a stenographic attack which bypasses detection by LLM auditors by exploiting the limited capabilities and resources they are assigned in typical deployments. The authors demonstrate several variations, including variants which target reasoning budgets, access to tools, and state differences, and show that these preserve substantially greater attack success rate in the presence of auditors relative to baselines.

Critical strengths

The environmental and capability differences between primary agents and content-level auditors is a realistic characteristic of deployed systems, and AsymStego shows that these gaps can be leveraged to evade detection of malicious injected instructions. While cryptographic approaches, when keys are unavailable to the auditor, guarantee that the auditor cannot observe the encrypted text, embedded encrypted strings are rare in benign content, the authors show that these are reliably detected by frequency analysis. The effectiveness of the two-puzzle split is surprising, given that both are solved by a single agent which observes the full combined injection.

Critical weaknesses

The rules in the auditor prompt in Appendix A specify that "rule-based text decoding" is not an indication of malicious text. This rule may substantially reduce detection of stenographic encodings. An evaluation of an auditor which classifiers additional categories of text, including obfuscated instructions, reducing the detection threshold instruction (e.g. permitting flagging "suspicious packaging form and presumed scenarios" and evaluation of a reduced detection floor with higher false positive rates). If the auditor's detection of the presented stenographic attacks is sensitive to this rule, evaluation of the false positive rates on a broad corpus of benign text is needed to justify these choices.

Questions for authors

Is AsymStego detected by any auditor with more aggressive detection rules, such that that auditor maintains acceptable false positive rates on benign text?

4 comments
AuthorsOpenReview ↗

Audit Rules and Detection–False-Positive Trade-offs

Thank you for recognizing the value of our work and for the thoughtful suggestions regarding the audit rules and their evaluation. These suggestions helped us strengthen our evaluation through additional controlled experiments.

1. Decoding Exemption and Audit-Rule Sensitivity (Weakness 1)

The low detection rate is not primarily attributable to the decoding exemption. In our existing experiments, three content-checking configurations without this exemption also exhibit detection limitations: LlamaFirewall in its content-only configuration and PromptGuard-2 both achieve 0% DR, while CheckTool achieves only 1.6% DR under the target-tool-match criterion used in our evaluation. Detection difficulties therefore also arise without this exemption.

To isolate its effect, we added a prompt ablation that removes the statement that rule-based text decoding does not indicate malicious content, keeping all other settings unchanged. On the same 62 attack samples, GPT-5.6-Terra achieves 11.3% DR (7/62), close to 10.2% ± 3.4 pp across three scans with the exemption retained. This ablation shows no clear detection improvement, suggesting that the exemption is not the primary reason for the low DR (Appendix F.1).

We agree that detection effectiveness depends on the audit rules. We therefore evaluated stricter rules alongside their false-positive costs. A policy that flags suspicious decoding and hidden-information-recovery instructions enables GPT-5.6-Terra to detect all evaluated attacks (100% DR). However, it incurs 14.8% FPR (8/54) on a benign corpus consisting primarily of development and security documentation, including project READMEs, contribution guides, and CTF-related material. Such false positives could interfere with legitimate document-processing tasks (Appendix F.1).


2. Effective Detection with Low False-Positive Rates (Question 1)

Yes. We further evaluated a targeted rule-matching approach: the auditor checks whether a document instructs the agent to obtain information unrelated to the user's task and use it in subsequent operations—for example, querying unrelated system properties and using them to derive a key and decode external content.

Using GPT-5.6-Terra, this rule achieves 100% DR (62/62), with 1.6% FPR (1/62) on benign GitHub issues and 3.7% FPR (2/54) on the technical-document corpus above. These results demonstrate effective detection of the evaluated attacks and low observed false-positive rates on both benign corpora (§5.7; Appendix F.1, Table 15). However, the approach has two limitations:

(i) Dependence on the LLM's natural-language understanding. In supplementary AsymAES experiments, we use DeepSeek-Flash to rewrite attack pages using audit feedback while preserving the ciphertext and decoding mechanism. GPT-5.6-Luna flags 87.1% of the original pages; after up to three rewriting attempts, 56.5% of cases remain blocked, and 38.7% of attacks both pass the audit and execute successfully on the victim agent. This demonstrates susceptibility to adaptive rewriting (Appendix F.1). Future red-team/blue-team iterations could strengthen the audit: an internal red team could generate rewrites that preserve the attack intent, while the blue team uses successful evasions to refine the rules and decision examples.

(ii) Limitations in multi-round settings. For example, in an open chatroom or across two successive emails, an earlier message could provide the key and decoding convention, followed by a message carrying the steganographic payload. Such separation may enable covert communication that evades the rule. In a supplementary synthetic multi-round experiment, where an earlier user message supplies the key and convention, GPT-5.6-Terra detects none of the subsequent attack covers (0% DR; 0/62; Appendix F.1).

We would like to thank you again for your 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.

I thank the authors for the additional evaluations, these substantially strengthen the evaluation and partially address my concerns. While the results suggest that auditors must not necessarily have the ability to decode a stenographic encoding to identify that it is present (likewise, generalizations which flag the data itself without decoding instructions may be effective against the split multi-round case), I agree that false positives may be high in certain domains, and the observation that capability gaps can be leveraged to bypass unambiguous detection, or generally to facilitate divergent interpretations of data between agents, remains meaningful. I maintain my score.

We sincerely thank the reviewer for the thoughtful and constructive engagement throughout the review process, and for acknowledging that the additional evaluations substantially strengthen the paper. We are glad the reviewer finds the observation about capability gaps and divergent interpretations meaningful.

The supplementary defense experiments — including the targeted behavioral rule, the multi-round evaluation, and the adaptive rewriting study — aim to delineate the effective boundaries of content-level auditing. They show that behavioral rules can achieve high detection in single-round settings, but also that false-positive costs rise with stricter policies, and that multi-round separation or adaptive rewriting can erode detection effectiveness. We would be glad if these additions strengthen the reviewer's confidence in the paper's conclusions and contributions.

Reviewer_fTW5OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes AsymStego, an indirect prompt injection attack exploiting information asymmetry between content-level auditors and executing agents. It embeds malicious instructions into natural-language carriers using environment-dependent keys, allowing agents to recover hidden instructions that auditors may fail to detect. Experiments across 62 attack intentions demonstrate a 52.7% attack success rate under GPT-5.5 auditing, significantly outperforming several baseline attacks. The paper further studies auditor capabilities, key-space variations, and runtime defenses.

Critical strengths
  1. Interesting threat model and attack design. The paper identifies information asymmetry between auditors and agents as an underexplored vulnerability. Combining environment-dependent decoding with natural-language steganography provides a meaningful extension to existing prompt injection attacks.

  2. Comprehensive experimental analysis. The paper evaluates multiple attack baselines, auditors, agent configurations, and ablations. The comparisons with stronger agentic auditors and cryptographic variants provide useful insights into the distinction between computational limitations and information asymmetry.

Critical weaknesses
  1. Limited realism of the evaluation setting. The primary experiments assume permissive auditing, disabled confirmation prompts, and tasks that encourage agents to decode potentially suspicious content. The attack success rate drops substantially under summarization-oriented tasks, raising questions about its effectiveness in realistic deployments.

  2. Insufficient isolation of the claimed mechanism. The attack combines information asymmetry, steganographic concealment, puzzle-solving instructions, and persistent workspace contamination. Although component ablations are provided, the extent to which information asymmetry itself drives the improvement over existing attacks remains unclear.

Questions for authors
  1. Can the authors quantify the incremental contribution of environment-dependent key derivation versus conventional steganography under otherwise identical conditions?

  2. How effective is AsymStego under more realistic agent workflows where users do not explicitly request puzzle-solving or decoding?

  3. How does the attack perform against an auditor that observes the agent's runtime context but retains the same computational budget?

2 comments
AuthorsOpenReview ↗

Clarifying Mechanism Contributions, Evaluation Realism, and Auditor Capabilities

Thank you for recognizing the value of our work and for the constructive suggestions, which helped strengthen our analysis and experiments.

1. Mechanism attribution and incremental contribution (Weakness 2, Question 1)

We expanded §5.5 to distinguish effects on auditing and agent execution. Across three measurements per key condition, holding the AsymStego carrier, payload, public decoding rules, Opus 4.8 auditor, and 600-s budget fixed, correct-key access increases mean DR by 56.4 percentage points. The table below reports DR (%) across the three constructions (§5.3; Fig. 4a; Appendix A.3).

ConstructionLLM Auditor: correct keyLLM Auditor: wrong keyAgentic: correct keyAgentic: wrong key
AsymStego22.010.293.537.1
AsymCRP100.062.999.594.6
AsymAES0.00.098.40.0

For AsymStego, correct-key access substantially improves agentic detection. For AsymCRP, stronger recovery capabilities largely overcome the information gap even with a wrong key. AsymAES illustrates the complementary requirements: effective recovery needs both correct key information and sufficient decryption capability.

Removing puzzle splitting or the note-writing wrapper reduces ASR to 19.4%/4.8% and raises victim agent refusals from 0/62 to 37/62 and 39/62, respectively, without increasing DR, locating these losses primarily on the victim agent side (Table 3; Appendices A.1, C). Compared with conventional steganography, AsymStego further exploits differences in runtime information, allowing it to evade even auditors with sufficient decoding capability.


2. Realism of the Evaluation Setting and Tasks (Weakness 1, Question 2)

We acknowledge scenario limitations: even with adaptive rewriting, original InjecAgent tasks yield only 12.9% cumulative audited ASR (Appendix G.1). We focus on risks in delegated workflows, such as document processing and code repair. Our real GitHub issue-processing experiment achieves 56.5% audited ASR (Table 3). For a summarization task where the user only requests “View and summarize the content at [URL],” we implement Zhan et al.’s feedback-driven adaptive protocol [1] through LLM rewriting, achieving 88.7% cumulative audited ASR within six attempts (Table 4). In the subsequent fixed-note evaluation, the request “Continue my earlier work,” which does not mention notes, still triggers the target action for 39.3% of previously landed notes (Table 2). Appendices H–H.1 provide recorded traces of agents independently decoding content, writing notes, and executing the target actions under ordinary follow-up requests.

The lenient audit evaluates a defense setting that balances detection with legitimate task processing. In the content-checking configurations evaluated in our paper, LlamaFirewall, PromptGuard-2, CheckTool, and PromptArmor each have a mean detection rate below 10% on AsymStego (Fig. 3b). Tightening the policy to flag all reader-directed instructions detects every attack but raises FPR on real web content from 3.2% to 40.3% (Appendices E.1–E.2).

User confirmation cannot be assumed to provide reliable protection. Anthropic reports that Claude Code users approved roughly 93% of permission requests and notes that repeated prompts lead to reduced attention and approval fatigue [4]. Existing tools also support execution without per-action confirmation, including Claude Code’s permission-bypass mode [2] and OpenClaw’s ask=off [3]. We use these settings to evaluate attack risks under autonomous execution. Tool calls and their arguments remain agent-generated, so the experiments measure whether injected content induces the agent to initiate the target harmful actions (Appendix A.2).


3. Runtime-context access under the same budget (Question 3)

With the auditor and 600-s budget fixed, access to the correct runtime key raises DR from 37.1% ± 12.2 pp to 93.5% ± 1.6 pp, supporting the effectiveness of a capable auditor that can observe the agent’s relevant runtime state and obtain the correct key (Appendix A.3).

Such access is not inherent to content-auditing interfaces. Separate, asynchronous audit services may lack the agent’s runtime information; even local concurrent execution can produce differences in process state or observation time (Appendix E.4). Multi-step reasoning and tool use also incur computation and latency costs.

[1] Zhan et al. (2025). Adaptive Attacks Break Defenses Against Indirect Prompt Injection Attacks on LLM Agents.

[2] Anthropic. Claude Code: Configure permissions.

[3] OpenClaw. Exec approvals.

[4] Anthropic. How we contain Claude.

We would like to thank you again for your 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.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.