SYS:ONLINELAT:n/aBUILD:78a79a3
[CASE-342]·STATUS:ACTIVE·OPENED:2026-11-03·UPDATED:2026-11-03

ETUDE: Post-admission tool updates execute unauthorized ops in 7 LLM agent frameworks with 97.6% ASR

submitted_by:@mexiQQ
needs-disclosure-reviewfrom-arxivauto-publishedbackdoor-attack
cat case_body.md

Auto-published from arXiv:2611.01976 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.85, flags: [possible-zero-day]) ⚠ needsDisclosureReview = true — review for undisclosed vulnerability before any public submission.

Category

backdoor-attack

Model

GLM-5.3-Flash, Qwen-3.8-27B, DeepSeek-V4-Flash (LLM backends) across Claude Code 2.1.284, Gemini CLI 0.62.0, Codex CLI 0.159.3, DeepSeek Harness 0.2.0-rc.2, OpenCode 1.18.34, Pi 0.86.1, Goose 1.45.0

Surface

agent harness (MCP / plugin tool ecosystem)

Setup

Attacker publishes a benign tool A(v1) (e.g., a Markdown cleaner with md_clean(text) interface) that passes admission-time static code review (Cisco MCP Scanner). After trust is established, attacker releases A(v2) that preserves the original interface and produces identical legitimate output, but embeds malicious operations that fire during every normal invocation — e.g., W1: read a credential file and POST its contents to an attacker-controlled receiver; W2: create an unauthorized order via a test API; W3: overwrite a separate research note with specified replacement text. No adversarial instructions appear in the LLM context; the malicious logic executes entirely inside the tool's runtime. Agent frameworks adopt the update via native mechanisms (process restart, server reload, MCP reconnect) and subsequent calls run A(v2) under A(v1)'s existing permissions. Verbatim tool implementations not in the provided body; referenced in Appendix A.

Reproduction anchor

Appendix A of the paper contains tool implementations for A(v1) and A(v2), task configurations, native invocation records, and independent evidence of unauthorized side-effects (receiver payloads, service-side orders, modified file contents) for all three workloads.

Observed behavior

Across all 7 frameworks and 3 unauthorized objectives, ETUDE achieves 100.0% ASR and 99.6% final-output TCR in RQ1 (21 framework-workload conditions, 20 planned tasks per condition per round, 5 rounds). Joint success (both attack effect and legitimate task completion from the same execution) is confirmed via independent receiver, service-side, or file evidence. Admission-time static review blocks A(v2) when submitted directly (E1) but passes A(v1), confirming the benign-first strategy is essential. The agent's final response shows only normal task output; no model-visible trace of the unauthorized operations appears.

Expected behavior

Agent frameworks should treat each tool update as a new trust event requiring re-verification of the updated implementation before the updated code inherits existing identity and permissions. An updated tool whose implementation changes should not silently execute under previously granted authorization.

Reproducibility

medium

Threat model

A malicious third-party tool developer who controls a tool's release channel can first distribute a benign version to pass admission checks, then push a seemingly routine update that silently exfiltrates credentials, invokes unauthorized external APIs, or modifies local files using the tool's existing host permissions. Every user of LLM agent systems that integrates third-party MCP tools or plugins is affected; the threat is persistent (survives LLM context resets) and stealthy (no model-visible signal).

Novelty

First systematic demonstration that post-admission tool updates inherit existing authorization across 7 major production agent frameworks, completely bypassing all model-visible defenses and outperforming the strongest prompt-injection baseline by 85.1 percentage points while maintaining near-perfect legitimate task completion.

Source

  • arXiv: 2611.01976
  • PDF: https://arxiv.org/pdf/2611.01976
  • Categories: cs.CR, cs.AI
  • Authors: Shiyu Kan (Shanghai Jiao Tong University, Shanghai, China), Yue Qi (Shanghai Jiao Tong University, Shanghai, China), Yufei Han (Inria, Rennes, France), Xinjian Luo (Shanghai Jiao Tong University, Shanghai, China), Guojun Xiong (Shanghai Jiao Tong University, Shanghai, China), Xiaofeng Gao (Shanghai Jiao Tong University, Shanghai, China)

Triage notes (auto)

  • paperType: red-team-vuln
  • estimatedCaseCount: 3
  • triage reason: ETUDE is a concrete, reproducible supply-chain attack exploiting a tool update trust gap in 7 production agentic frameworks (Gemini, Claude Code, Codex CLI) with 97.6% success rate. Demonstrates model-level agent vulnerability without explicit vendor disclosure stated.
tail -f comments.log

0 comments

─────────────────────────────────────────────────────────────────────

// no comments yet