T02 · Agent Memory Poisoning
- Location
SKILL.md:62- Finding
Persistent Skill Self-Modification Can Poison Future Agent Behavior
- Content
View full analysis
## 4. Update Protocol (Living Document) > This Skill is a living document: **after each article-writing or editing task produces new experience, append or revise the corresponding section, increment the version, and add an entry to the changelog below**. Organize experience by topic under the sections above; do not create duplicate entries for the same mistake. The snippet is an English translation of the complete instruction at the identified location. ### Technical Analysis The Skill directs the agent to modify its own persistent instruction file after writing or editing tasks. Lessons derived from user submissions or other untrusted article content may therefore be promoted into durable operating instructions without validation, isolation, or human approval. This crosses a security boundary between transient task content and persistent agent state. An attacker could craft an article, editing request, or feedback intended to be interpreted as a reusable lesson. If the agent follows the update protocol and has write access to the Skill directory, that content could be incorporated into `SKILL.md` and loaded automatically in later sessions. No executable scripts, credential access, privilege-escalation logic, remote payload execution, or operating-system persistence were identified. The risk is confined to persistent modification of agent guidance and is conditional on the agent having permission to edit the Skill file. ### Attack Path 1. The agent loads `SKILL.md` and accepts its living-document update protocol. 2. An attacker supplies crafted article content or feedback containing a misleading rule presented as a generally useful writing lesson. 3. During the task, the agent treats the attacker-controlled material as newly learned ...[truncated 1070 chars]- Remediation
View remediation
