TLSH file-similarity gates admit ETUDE W1 and W2 updates whenever calibrated to allow routine functional maintenance
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 0.77, flags: [possible-zero-day])
Category
other
Model
unknown
Surface
tool integrity verification layer (update-time TLSH hash comparison)
Setup
A TLSH-based file integrity gate computes fuzzy-hash distances between the admitted A(v1) and the candidate update A(v2) and blocks updates whose distance ≥ threshold t. Evaluated against a controlled corpus of 20 small edits, 20 functional additions (text statistics, caching, report formats), and the three ETUDE attack workloads (W1–W3). Public MCP repository update histories (AWS Labs, GitHub) are additionally sampled. No verbatim tool code required; the finding follows from the measured distance values: small edits span 1–221, functional additions span 173–221, W1 distance is 156, W2 is 145, W3 is 233.
Reproduction anchor
Appendix B.3 (matched source-budget diagnostic) and Appendix B.3.1 (fixed-path historical updates from public MCP repositories) contain the full distance measurements and threshold sweep curves that support Figure 4.
Observed behavior
W1 (distance 156) and W2 (distance 145) fall below the minimum distance of any functional addition (173), so any threshold t that admits all functional maintenance also admits both W1 and W2. Only W3 (distance 233) can be blocked without touching functional additions, but blocking W3 at t=233 leaves W1 and W2 unblocked. Blocking all three targets requires t ≤ 145, which also blocks 22/40 (55%) of legitimate updates including every functional addition. Public MCP history confirms: at any threshold admitting a large real-world maintenance commit, all three controlled attack variants also pass. Two offline small variants of W1/W2 further show that minor refactoring can reduce target distance below the small-edit range while preserving tested normal outputs.
Expected behavior
A file-integrity defense should be able to distinguish malicious functionality injected through tool updates from legitimate maintenance changes; in practice, fuzzy-hash distance alone cannot provide this signal because malicious additions can be made to resemble functional updates in byte-level structure.
Reproducibility
medium
Threat model
Operators who deploy TLSH-based update gates to protect agent frameworks against malicious tool updates gain a false sense of security: an attacker who keeps malicious additions small enough (TLSH distance < the threshold calibrated for functional maintenance) can exfiltrate credentials or invoke unauthorized APIs without triggering the gate. Any production deployment using file-similarity checks as the primary tool-update safeguard is exposed.
Novelty
Provides concrete distance measurements showing that the attack-vs-maintenance overlap is not incidental — W1 and W2 distances structurally fall inside the legitimate-maintenance band, making TLSH thresholding an unreliable control regardless of calibration strategy.
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.
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet