learn
Pass
Audited by Gen Agent Trust Hub on Nov 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONPERSISTENCE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to ingest user-provided 'lessons' and persist them into instructions files (e.g., CLAUDE.md, lessons.md, and skill overlays) that are automatically loaded into the agent's context in future sessions. This creates a significant surface for persistent indirect prompt injection.
- Ingestion points: User input provided via the $learn command or inferred from conversational triggers (SKILL.md).
- Boundary markers: The skill instructs the agent to apply several 'Quality Gates' (Lesson Quality Gate, Learned Guidance Quality Gate) to sanitize and generalize input, but these are natural-language instructions and not robust security boundaries.
- Capability inventory: The skill has the capability to modify critical project files including CLAUDE.md, project-config.json, and various reference documents, and can trigger synchronization scripts across mirrors.
- Sanitization: The protocol mandates a 'GENERALIZE FIRST' rule to strip specific incident nouns (files, tools, domain names), which serves as a form of instruction-based sanitization.
- [PERSISTENCE]: The skill provides a mechanism for maintaining state (in the form of rules and context) across sessions by modifying documentation and configuration files in the repository. While this is the core intended feature, it represents a persistence mechanism that could be leveraged to maintain malicious instructions or configuration states if the agent is compromised or deceived.
- [COMMAND_EXECUTION]: The skill executes local shell commands and scripts (e.g., node .claude/hooks/lib/project-config-schema.cjs, node .claude/scripts/build-protocol-projection.cjs, and node .claude/skills/sync-codex/scripts/run-codex-sync.mjs) to manage configuration and synchronization. These scripts reside within the project's own configuration directory (.claude/).
Audit Metadata