Shanghai Jiao Tong University and Inria propose the tool-update attack ETUDE, with a 97.6% success rate across configurations in controlled experiments
Agent的工具更新执行攻击
Researchers from Shanghai Jiao Tong University and Inria propose ETUDE, a tool-update direct-execution attack that does not manipulate model decisions but instead exploits trust that is not invalidated in sync after a tool update, with malicious operations executed by the tool's own code. In experiments on seven agent frameworks including Gemini CLI, Codex CLI, and Claude Code, the paper reports an overall attack success rate of 97.6% and a normal task completion rate of 95.9%. The attack premise is that the attacker can control the updated tool code, the update is adopted by the execution environment, and the tool has the corresponding permissions; the experiments were conducted in a controlled environment built by the researchers. Defense tests show that a strict integrity threshold would falsely block 55% of normal updates, that after the attack code is obfuscated with PyArmor ETUDE still has an 85.0% success rate under static review, and that runtime auditing can raise alerts but has cost and coverage issues. The authors recommend binding tool authorization to code versions, minimizing tool execution permissions, and establishing joint auditing of model behavior and tool execution behavior.
The article reviews an attack path that bypasses model decision-making and directly executes malicious code via tool updates, along with the limitations of three types of defenses; teams working on agent security can use it to check their tool-update trust management.