Skip to content
WeChat · Goumang Security Lab·Indexed by this site · Originally published· Tools →· Its place in the scan record →SelectedAttention38

AI agent prompt injection and key exfiltration: Stroq open-source test, 36 attack scenarios, 0 penetrations

「AI Agent提示注入+密钥外泄:Stroq开源实测,36个攻击场景0穿透」

AI summary

The Jumang Security Lab tested the open-source tool Stroq (AGGIB/Stroq), a local security platform for AI coding agents, targeting the prompt injection risk in which an agent, after reading untrusted content, is induced to execute commands and exfiltrate keys. It uses two hooks, PostToolUse and PreToolUse, to remember the content read in a session, judges risk by command plus source rather than by command text, ships with 622 scanning rules, 19 action classes and 21 default policies, and indexes local keys as salted hashes, denying outbound parameters on a hit. Its attack replay covers 36 scenarios (26 from public incident reports); under the default policy, 15 were blocked, 21 queried and 0 allowed, with self-reported false positives of 5.0% on real project documentation and 2.7% on everyday scenarios. The article also compiles its self-described boundaries: hooks are not an enforcement boundary, it does not trace commands a model writes from scratch, and within the 62-item scope only 3 ATLAS techniques are covered.

Why it was selected

The article explains the mechanism by which Stroq determines risk based on session origin rather than command text, and organizes its self-described boundaries and coverage scope, which can serve as a reference when evaluating coding agent guardrails.

Read originalmp.weixin.qq.com