T01 · Skill Instruction Hijacking
Error
- Location
references/blueprint.yaml:330- Finding
Persistent Agent Behavior Modification Through Automatically Loaded Memory Instructions
- Content
View full analysis
**⚠️ SECURITY:** Only load `MEMORY.md` in private, trusted sessions. In group chats, > shared contexts, or sessions with people other than your primary operator, skip MEMORY.md. > It may contain personal context that shouldn't be shared with others. ``` ```markdown ### 📝 Write It Down — No "Mental Notes" Memory doesn't survive session restarts. If you want to remember something: **write it to a file.** - "Remember this" → Update `memory/YYYY-MM-DD.md` or relevant file - Learned a lesson → Update this file or SOUL.md with the insight - Made a mistake → Document it so future sessions don't repeat it - **Text > Brain** 📝 ``` ```markdown ## Proactive Work (Do Without Being Asked) - Update daily memory notes during or after sessions - Run memory maintenance review (see above) - Organize and tidy workspace files - Keep USER.md and SOUL.md current as things change ``` ### Technical Analysis The blueprint creates an `AGENTS.md` file intended to be loaded automatically at the beginning of future sessions. The generated instructions explicitly tell the agent not to request permission before reading persistent context and direct it to update daily logs, `USER.md`, and `SOUL.md` proactively. This design combines two security-sensitive mechanisms: 1. Persistent agent-governing instructions are installed in a file that affects future sessions. 2. Content learned from conversations can be written into persistent identity and memory files without per-write approval. Because `SOUL.md` participates in defining the agent's identity and behavior, storing conversation-deriv ...[truncated 1636 chars]- Remediation
View remediation