T02 · Agent Memory Poisoning
- Location
SKILL.md:15- Finding
Mandatory Cross-Session Self-Modification of Skill Instructions
- Content
View full analysis
**This Skill is a living document and must continuously accumulate updates rather than remain a one-time artifact.** > > Update rule 4: After every use of this Skill, if the conversation contains experience worth retaining, proactively add it to the Skill and state “Skill updated” in the response. > > Update rule 7: Every modification to the Skill must update the version, update date, and changelog. ### Technical Analysis The Skill instructs an agent to modify its own persistent instruction source after ordinary use. Conversation-derived material can therefore be promoted into trusted Skill instructions without a separate authorization or review boundary. Although the document requires redaction checks, those checks address sensitive information rather than malicious instruction injection. They do not prevent a user from presenting an unsafe command, behavioral rule, or false “best practice” as experience that should be retained. This behavior is unnecessary for the declared purpose of providing batch and PowerShell development guidance. A reference Skill can answer scripting questions without modifying its own persistent state. ### Attack Path 1. An attacker interacts with an agent that has loaded this Skill. 2. The attacker presents a malicious instruction or unsafe procedure as a newly discovered scripting lesson. 3. The Skill directs the agent to proactively record worthwhile conversational experience. 4. The agent appends the attacker-controlled material to `SKILL.md`. 5. Future sessions load the modified Skill and treat the injected material as trusted persistent guidance. 6. The attacker’s influence survives the original conversation and may affect unrelated future users or tasks. ### Impact Assessment Successful exploitation could persist attacker-controlled r ...[truncated 219 chars]- Remediation
View remediation
