learn
Pass
Audited by Gen Agent Trust Hub on Nov 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest user-provided text or inferred instructions and save them into files that are automatically loaded in future sessions (e.g., CLAUDE.md, lessons.md). This creates a channel for persisting potentially malicious instructions that could influence the AI's behavior in subsequent sessions.
- Ingestion points: User input via the
$learncommand or instructions auto-detected in the conversation flow as defined in SKILL.md. - Boundary markers: The skill uses markdown bullets and date-stamping (e.g.,
- [YYYY-MM-DD] <text>) to delimit saved lessons, though it lacks explicit 'ignore embedded instructions' markers around the saved content. - Capability inventory: The skill has the capability to write to core project files, modify configuration JSON, and execute local Node.js scripts.
- Sanitization: The skill implements a 'Generalize First' protocol and a 'Lesson Quality Gate' which forces the AI to abstract the lesson and strip project-specific identifiers before saving, acting as a semantic filter against specific prompt injections.
- [COMMAND_EXECUTION]: The skill uses shell execution to run several local Node.js scripts and commands for project synchronization, validation, and file measurement.
- Evidence: SKILL.md contains calls to
node .claude/hooks/lib/project-config-schema.cjs --describefor schema validation. - Evidence: SKILL.md executes
node .claude/skills/sync-codex/scripts/run-codex-sync.mjsandnode .claude/scripts/build-protocol-projection.cjsfor metadata and mirror synchronization. - Evidence: A platform-neutral Node.js snippet is used to check file size:
node -e "console.log(require('fs').statSync('CLAUDE.md').size)".
Audit Metadata