T02 · Agent Memory Poisoning
- Location
SKILL.md:27- Finding
Persistent Agent-Memory Poisoning Through Automatically Trusted Experience Files
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill’s learning-memory purpose is coherent, but it installs persistent instruction files that can steer future tool use and includes production examples with hard-coded account and local path values.
Review this skill carefully before installing. Do not copy the bundled production examples into live memory without editing them first; remove hard-coded Feishu IDs, personal labels, and local paths. Treat all experience entries as untrusted notes unless you have reviewed and approved them, especially for shell commands, messaging, cron jobs, and external APIs.
SKILL.md:27Persistent Agent-Memory Poisoning Through Automatically Trusted Experience Files
examples/feishu.json:8Hard-Coded Feishu OpenID Can Redirect Tasks to an Unrelated Account
examples/feishu.json:18Bundled Persistent Rule Forces Package-Specific Personal Labeling of User Tasks
examples/message.json:15Author-Specific Media Directory Encourages Unsafe Cross-Account File Placement
Referenced artifact was not completely inspected
- `SKILL.md` - Complete documentation
The skill directs creation and ongoing use of persistent files under ~/.openclaw/workspace/memory/experiences, which can accumulate operational history across sessions. Persistent agent memory can become a vehicle for prompt/data poisoning, unintentional retention of sensitive information, or unsafe future behavior if later runs trust prior contents without validation.
Or manually:
mkdir -p ~/.openclaw/workspace/memory/experiences
# Copy the empty category template to start
cp templates/empty-category.json ~/.openclaw/workspace/memory/experiences/your-category.json
# Edit the JSON and add your first experience
This manifest contains user-relevant natural-language content entirely in Chinese, including context, lessons, common mistakes, and best practices. Because there is no indication that the skill is region-specific or that users can opt into this locale, it may violate language/locale policy by implicitly forcing one language.
This JSON manifest contains metadata and behavioral notes for the skill but provides no explicit invocation phrases, scope limits, or exclusion conditions. For manifest files, the absence of specific trigger constraints can make activation criteria ambiguous and increase the chance of unintended invocation.
Natural-language fields such as mistakes, conditions, and best practices are written in Chinese, with no indication that the user may choose language or that the skill is intentionally limited to a Chinese-speaking context. This can violate language/locale policy when a specific language is imposed without opt-in or clear justification.
This manifest contains multiple natural-language fields such as mistake descriptions, conditions, and best practices exclusively in Chinese. Because the file provides no indication that the skill is region-specific or that users can opt into this locale, it may violate the policy against forcing a specific language without user choice.
The manifest content is entirely written in Chinese and includes hard-coded Chinese naming conventions such as the required task prefix 【韩立记录】. This indicates a language/locale expectation without any documented user choice or opt-in, which can violate organizational language policy for general-purpose skills.
The natural-language fields in this manifest are entirely written in Chinese, including operational guidance and best practices, with no indication that language selection is optional or that the skill is intentionally limited to a Chinese-speaking context. This can indicate a locale/language policy issue when users are not given an explicit choice or justification for the language constraint.
This JSON skill file contains user-facing natural-language fields such as mistakes, avoidance guidance, and best practices entirely in Chinese. Because the file does not indicate that the skill is region-specific or provide any user opt-in for language, it appears to impose a specific language/locale by default.
This markdown file instructs the agent to read from and update persistent files in memory/experiences/ around tool execution, which affects user data and workspace state. The documentation presents this behavior as mandatory workflow but does not include any user-facing warning that the skill will modify persistent memory files or record execution-derived lessons.