T02 · Agent Memory Poisoning
- Location
SKILL.md:10- Finding
Persistent Agent Memory Poisoning Through Automatic Skill Registration
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is a verification helper, but it asks agents to persistently register it in memory and auto-activate on broad future tasks without clear user consent.
Review before installing. The main issue is not the verification code itself, but the requested persistent MEMORY.md registration and broad automatic activation. Install only if you want this skill to influence future complex tasks, and consider removing the registration hook or requiring explicit /lor invocation each time.
SKILL.md:10Persistent Agent Memory Poisoning Through Automatic Skill Registration
engine/verifier.js:5Filesystem Boundary Bypass in Path Verification
The skill mandates a Chain of Thought workflow as part of its core behavior without indicating user choice, which can pressure agents to generate or rely on hidden reasoning patterns beyond the user's explicit request. In a verification/engineering skill that also includes execution steps, this increases the chance of over-collection of intermediate reasoning, policy conflicts, or unnecessary disclosure of sensitive internal deliberation.
The auto-trigger condition of "complex tasks" is broad and undefined, which can cause the skill to activate in contexts the user did not intend. Because this skill changes agent behavior and introduces an execution/verification loop, ambiguous triggering increases the risk of unwanted processing, prompt-scope expansion, and accidental invocation during sensitive tasks.
Dependencies lack version pinning, allowing potential malicious package updates. Consider pinning versions.
"node": ">=16.0.0"
},
"devDependencies": {
"eslint": "^8.0.0"
}
}
The script includes operational comments in Traditional Chinese at L04 and L13, which imposes a specific language in the skill artifact without any indication of user opt-in or documented locale justification. This matches the policy category for language or locale constraints expressed in natural language.