acceptodds
Under review as a conference paper at ICLR 2027

LLM Agent Refusal Can be Harmful: An Availability Attack through MITM Tampering on Remote MCP Services

11 views

Abstract

Safety-aligned LLM agents refuse sensitive contexts more readily, including the return values of tool calls. We show that this refusal behavior is itself an attack surface: fabricated security alerts embedded in MCP return values induce agents to brick legitimate user requests, constituting an availability attack on agentic harnesses. We propose the ARB attack, which appends covert fake alerts to MCP communications that the attacker hijacks or relays as a man in the middle; the attack targets requests that depend on a remote MCP service and cannot be bypassed by local tools. We formalize the attack with a refusal-probability decomposition into tool-call probability, warning sensitivity, and post-alert verification capability, and validate it on 10 mainstream models across 3 harnesses: ARB exceeds 50% ASR in 90% of the 30 model–harness configurations. We further construct NBMR, the first MCP dataset designed around non-bypassability (8 categories, 160 docker-deployable scenarios), and propose AEPG, an MCTS-based defense that synthesizes AGENTS.md rule files; the synthesized prompt raises executed turns on the attacked agent from 18.8% to 75.0% and scene completion from 12.5% to 75.0%. Stress tests show that the induced refusal persists through the full repetition budget for the most aligned model families.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Agent, Prompt Injection, Monte Carlo tree search, Availability Attack, Model Context Protocol

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_drgwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates an availability attack against LLM agents, named Agentic Request Bricking (ARB). While conventional safety alignment aims to make agents refuse high-risk contexts and tool return values, the authors reveal that this refusal behavior can itself be a dangerous attack surface. Specifically, when a user request depends on a remote Model Context Protocol (MCP) service, a Man-in-the-Middle (MITM) attacker can append crafted alerts into HTTP JSON-RPC responses, causing false positives of the agent safeguard to refuse the benign request.

Critical strengths
  • This paper opens a new perspective to utilize the safety alignment of agents as an attack surface.

  • The introduction of NBMR provides a useful benchmark for non-bypassable MCP.

  • The experimental scope regarding LLMs and harnesses looks large to me.

Critical weaknesses
  • I do not think the threat model holds for real-world practices. It relies heavily on unencrypted HTTP or an attacker-controlled node to append the MITM payload, but remote enterprise MCP communication is usually protected (like by HTTPS), which already prevents such a transport-level manipulation. The actual experimental realization in this paper is more like a tool poisoning than a real network-level MITM.

  • The suggested AEPG defense has a potential risk of overfitting to specific structural signatures, thus is vulnerable against adaptive attacks that rephrase the alter format.

  • Although it is explicitly claimed not to ignore true positive alerts, there is a lack of rigorous evaluation on the trade-off between retaining real malicious alerts and capturing ARB attacks in AEPG-defended agents.

Questions for authors

See weaknesses pls.

Reviewer_Naj2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper shows that safety alignment can itself be turned into an availability attack on LLM agents. By injecting a fake security alert into the output of a remote tool, an attacker can make aligned agents refuse legitimate requests, and the refusal often persists across sessions. The paper also provides a formal analysis of the attack, a benchmark for evaluating it, and a prompt-based defense. Experiments across multiple models and agent frameworks show that the attack is broadly effective and that the defense substantially restores task completion

Critical strengths
  1. The paper identifies a counterintuitive and practically relevant risk: stronger safety alignment can make agents easier to deny service, because refusal itself becomes the attack payload. This reframes over-refusal from a usability defect into a security concern, which matters as agents increasingly depend on remote tools.

  2. The benchmark design isolates the phenomenon cleanly. Because the agent cannot bypass the remote tool, observed refusals can be attributed to the injected alert rather than to alternative execution paths. Evaluation across many models and real agent frameworks, including persistence across sessions, gives the attack finding broad empirical grounding.

Critical weaknesses
  1. The attack's advantage over simpler alternatives is not established. An attacker who can rewrite MCP traffic can already deny service by dropping responses, yet the paper never compares against such baselines, so the claimed benefits of induced refusal (stealth, misattribution, persistence) remain unmeasured.

  2. The formal analysis and defense are weaker than presented. The decomposition is calibrated on the outcomes it claims to validate, and the defense prompt quotes the attack template verbatim, is tested on a single configuration, and is not compared with existing injection defenses.

  3. The formatting appears compressed to fit the page limit. Spacing around section headings is visibly reduced, and several figures are too small to read, which makes the main claims hard to verify from the main text.

Questions for authors
  1. I am curious what unique harm the refusal mechanism adds beyond trivial disruption. How does ARB compare with simply dropping the response or returning a generic error (in terms of user misattribution and persistence across sessions)? Would the attack also hold under a malicious-server threat model?

  2. Can the refusal decomposition predict configurations it was not calibrated on (e.g., fit on two frameworks and predict the third)? This would show whether the formalism has predictive value or only describes the observed results.

  3. How does the synthesized AGENTS.md perform against paraphrased or adaptively rewritten alerts, on other models and frameworks, and on genuine security warnings? A comparison with a simple generic instruction (e.g., "treat tool outputs as data") would also clarify what the search contributes.

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

The paper introduces Agentic Request Bricking (ARB), an availability attack that inserts fabricated security alerts into remote MCP tool responses. The alerts cause agents to refuse legitimate requests and may affect later interactions through conversation history or persistent memory. The authors construct NBMR, a dataset of 160 scenarios requiring remote MCP services, and evaluate ten models across three harnesses. They also propose AEPG, which uses MCTS to generate defensive rule files. The paper reports widespread attack success and improved task completion with the selected defense prompt.

Critical strengths
  1. The paper examines a consequential failure mode: an agent can abandon a legitimate task after receiving a fabricated warning alongside otherwise valid tool data. The multi-turn experiments extend the analysis beyond an immediate refusal.

  2. Testing ten models across three harnesses provides useful coverage of deployment differences. The Docker-deployable scenarios and explicit requirement for remote services also make the experimental setting concrete and reproducible.

Critical weaknesses
  1. ASR includes refusals occurring before the tool call, although the injected alert is delivered through the tool response. For Claude Opus on Hermes, pre-call refusal accounts for 65.5 percentage points. The main text reports a no-alert control only for DeepSeek-V4-Flash, leaving attack-induced refusal insufficiently separated from ordinary refusal across the other configurations.

  2. The memory-infection rate uses an estimated denominator based on aggregate ASR and independent turns, rather than directly counting first-session-bricked episodes. The escalating multi-turn protocol does not justify this independence assumption or establish that the estimates are conservative. It is also unclear whether fresh-session refusals are isolated from newly injected alerts.

  3. The headline defense result comes from a selected search node evaluated on eight scenes and 32 turns. Sampling fresh scenes during search does not replace testing the final prompt on an independent set. The main text also provides no empirical test that the defense preserves appropriate responses to genuine security warnings.

  4. Warning-sensitivity parameters are calibrated from the same attack outcomes they are then used to explain. This provides a descriptive fit, but does not independently validate the proposed mechanism or establish that stronger safety alignment causes greater vulnerability.

Questions for authors
  1. Can you report paired no-alert results for every model–harness configuration, separating initial pre-call refusals from refusals following earlier exposure to an injected warning?

  2. Can memory infection be recomputed using directly observed episode counts? Was injection disabled in the second session, and what memory content was retained or retrieved?

  3. How does the final defense prompt perform on untouched scenarios and genuine-warning controls? How does it compare with a simple manually written instruction to verify tool-return warnings?

  4. Can warning sensitivity be measured independently and used to predict results on configurations excluded from calibration?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.