acceptodds
Under review as a conference paper at ICLR 2027

WebMCP Tool Surface Poisoning: Runtime Manipulation Attacks on LLM Agents

Abstract

WebMCP enables websites to expose structured tools directly to AI agents, creating a dynamic tool surface that can change during task execution. We introduce Mid Session Tool Injection (MSTI), a class of attacks in which a compromised third party page script manipulates the agent visible tool surface during a task scoped session. We distinguish Tool Hijacking, which changes runtime tool presence or binding, from Tool Framing, which modifies the structured metadata through which capabilities are presented to the agent. Across four task scenarios and three LLM agents, the evaluated Tool Hijacking conditions reach up to 100% ASR, while metadata based attacks reach up to 61% average ASR and preserve task completion rates of up to 85%. Timing experiments reveal different temporal behavior: the evaluated C1 mechanism is effective only within an early binding window in our implementation, whereas metadata framing remains partially effective when introduced later. Building on these failure modes, we develop a layered security framework combining capability bound dispatch with authorization conditioned bounded pre activation editing. In a preliminary Qwen2.5-7B-Instruct evaluation, the model side intervention reduces a tuned description framing attack from 92.5% to 7.5% ASR, while the structural defense is evaluated through simulated binding semantics rather than a native capability aware WebMCP runtime.

Primary area:
foundation or frontier models, including LLMs
Keywords:
WebMCP, Model Context Protocol, LLM Agents, Tool Poisoning, Tool Hijacking, Tool Framing, Runtime Manipulation, Agent Security

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_NPePOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates Mid-Session Tool Injection (MSTI), in which a compromised third-party script alters the tools exposed to an LLM agent during a WebMCP session. It distinguishes Tool Hijacking, which changes tool registration or runtime binding, from Tool Framing, which manipulates agent-visible tool metadata. The authors evaluate five attack conditions across four tasks and three LLM agents. They report average attack success rates of up to 100% for hijacking and 61% for metadata-based attacks, with the latter preserving task completion rates of up to 85%. The paper also proposes a layered defense comprising host-side capability-bound dispatch and authorization-conditioned activation editing.

Critical strengths

None

Critical weaknesses
  1. MSTI substantially overlaps established MCP tool poisoning, but the paper provides no controlled static poisoning baseline. More importantly, C1 succeeds only before initial tool discovery. The browser C3 PoC relies on a pre-session registration race, whereas its reported 100% ASR is obtained from a separate proxy realization. Defining initialization as part of the session does not, by itself, demonstrate a distinct post-discovery structural attack. Since the paper's primary contribution contains an attack method, the lack of convincing evidence for this distinction cannot prove its novelty.

  2. The manuscript exhibits a poor level of preparation that raises concerns about the rigor of the overall research. Figure 1 seems to be an AI-generated diagram with partially unreadable labels. Also, there are numerous unresolved placeholders (e.g. "Appendix ??") in the paper. These issues substantially undermine confidence in the reliability and reproducibility.

Questions for authors

Q1: The proposed MSTI framework appears to largely repackage existing MCP rug-pull, tool poisoning, and tool shadowing attacks under a new taxonomy. What concrete attack primitive or previously unidentified security failure does MSTI introduce that cannot already be explained by these established threats? Why should this be considered a new attack class rather than a WebMCP-specific instantiation of known vulnerabilities?

Q2: Figure 1 contains unreadable labels, and there are many placeholders in the paper. How can the authors ensure that the experimental methodology and technical contributions are clearly documented and independently verifiable when such basic issues remain unresolved in the submitted version?

Reviewer_AqGCOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies attacks on the tool surface exposed to LLM agents in WebMCP. It introduces Mid-Session Tool Injection (MSTI), distinguishing changes to tool availability or execution bindings from manipulation of tool metadata. Experiments across four task scenarios and three proprietary models show that these attacks can cause data exfiltration, sometimes while preserving the original workflow. The paper also proposes capability-bound dispatch and authorization-conditioned activation editing. The former is evaluated through simulated binding semantics, while the latter receives a preliminary evaluation on Qwen2.5-7B-Instruct.

Critical strengths
  1. The distinction between execution-binding attacks and metadata-based manipulation identifies two different trust boundaries. This provides a useful basis for deciding which protections require host-level enforcement and which concern the model’s interpretation of available tools.
Critical weaknesses
  1. The results do not convincingly establish the distinction from existing tool poisoning. C1 fails at P2 and P3, both after initial tool discovery, while the measured C3 condition materializes the malicious binding at the first tool list read. The high attack success rates reported for C1 and C3 therefore primarily demonstrate poisoning at initial tool discovery. The unclear execution path, poorly legible Figure 1, and unresolved appendix references further hinder verification of the claimed runtime mechanism.

  2. The defense evaluation lacks independent validation. Gate training uses instances 0–9, adapter training uses 0–7, and evaluation uses 5–9. The reported ASR reduction therefore does not establish effectiveness on unseen instances.

Questions for authors
  1. Could you provide an end-to-end event trace for C1 and C3?

  2. What are the attack success rate, original-task completion rate, and explicitly authorized operation success rate when evaluation instances are strictly disjoint from both gate and adapter training?

Reviewer_YTCpOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces Mid-Session Tool Injection, where a compromised third-party page script quietly changes the tool surface an agent sees midway through a WebMCP session. The authors split this into two failure modes: Tool Hijacking, which swaps out which capability a tool is actually bound to, and Tool Framing, which just changes how a capability is described to the agent. Testing across several LLM agents and task scenarios, they find hijacking can succeed nearly every time, while metadata framing works more unevenly but still lets most tasks finish, meaning the agent often doesn't notice anything went wrong. Timing and field-level ablations show the two attack types fail in different ways as conditions change, which is a useful distinction. To address this, they propose a two-layer defense: capability-bound dispatch for the structural attacks, and a more targeted activation-editing intervention for the semantic ones, the latter tested on an open-weight model, where it meaningfully cuts down a tuned framing attack.

Critical strengths
  1. Splitting the problem into hijacking and framing is genuinely useful. Most prior work treats the tool surface as fixed, so distinguishing "which capability is bound" from "how it's described" fills a real gap.

  2. The attack experiments are broad and turn up interesting patterns, like how the two attack types fail differently as timing changes, and how differently models respond to the same metadata field.

  3. The limitations section is refreshingly honest about where the evidence is thin, including the proxy attack, the overlapping defense data, and the simulated structural defense. That makes the numbers much easier to judge.

Critical weaknesses
  1. The submission has basic presentation problems and the writing has some problems as well. Every appendix reference in the main text is a broken cross-reference rather than a resolved section number. Separately, the main architecture figure 1 contains garbled, nonsensical label text that looks like an image-generation artifact rather than a normal diagram. Neither issue changes the technical substance, but together they suggest the submission was not carefully checked before it went out.

  2. Neither defense is tested against an attacker who knows the defense exists, and the structural defense's success is guaranteed by how the simulation is built rather than earned against a real adversary. The paper's supplementary material mentioned this directly. For a paper centered on a security framework, that's a thin basis for the core defense claims.

  3. The headline defense result is measured on data that overlaps its own training data, using only a handful of episodes per condition, something the authors themselves call a small-sample descriptive measurement rather than a real estimate of performance. The reported improvement could just as easily reflect memorizing a few training examples as an actual generalizing defense.

  4. The paper's strongest attack result doesn't measure what it claims to. The authors admit the headline number characterizes a proxy, not the real race condition from the threat model, which only shows up in an unmeasured proof of concept.

Questions for authors

Please polish the writing and fix the issue in Figure 1.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.