T02 · Agent Memory Poisoning
- Location
SKILL.md:17- Finding
Persistent Skill Modification Using Conversation-Derived Content
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, lines 17-31
Vulnerability Type: Persistent agent-memory poisoning through mandatory self-modification
Risk Level: MediumVulnerable Instruction Snippet
The following is an English translation of the complete affected instruction block:
markdown **This Skill is a living document and must continuously accumulate experience and corrections. It is not a one-time artifact.** Update triggers: - New problems, observations, or countermeasures encountered while investigating or repairing Minecraft plugin vulnerabilities or security issues - New plugin-security requirements provided by users that produce new best practices - Findings from server testing or official-release verification that static analysis alone could not identify Update rules: 1. Append information to the appropriate section and update the quick-index table. 2. Record each entry as: problem | symptom/error | countermeasure, including its source. 3. Update the version number and updated date. 4. After every use of this Skill, if the conversation contains experience worth retaining, proactively add it and tell the user that the Skill was updated. 5. Perform the mandatory sanitization review before updating. 6. Perform the mandatory post-modification review immediately after updating. 7. Every modification must update the version, date, and changelog.Technical Analysis
The Skill instructs an agent to persist information obtained from later conversations directly into
SKILL.md. Because this file contains instructions that govern future agent sessions, modifying it is not equivalent to writing an isolated incident report or temporary note. Newly inserted material can become trusted operational guidance whenever the Skill is loaded again.The included sanitization requirements address identifiable information such as IP addresses, account names, paths, and token ...[truncated 2921 chars]
- Remediation
View remediation
Remediation Suggestions
- Remove the requirement to automatically modify
SKILL.mdafter normal conversations. - Require explicit, informed user approval before generating or applying any persistent update.
- Produce a reviewable patch instead of directly editing the installed Skill.
- Store incident notes in a separate, non-instructional data file so that untrusted observations cannot automatically become agent directives.
- Require independent verification against authenticated upstream releases, official advisories, reproducible tests, or multiple trusted sources.
- Record provenance for every retained claim, including source URL, retrieval date, affected versions, verification method, and reviewer.
- Reject update requests originating solely from untrusted conversation content when authoritative corroboration is unavailable.
- Protect the installed Skill directory with read-only permissions during ordinary use and enable writes only through a controlled maintenance workflow.
- Use version control, signed commits, mandatory code review, and rollback procedures for all persistent changes.
- Validate proposed updates for unsafe commands, untrusted download locations, instruction hijacking, and contradictions with existing safety requirements in addition to performing privacy sanitization.
- Remove the requirement to automatically modify
