The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill maintains a local database of penetration testing patterns that are injected into the agent's context during the reconnaissance and weaponization phases. This creates a surface where findings from one target could influence the agent's behavior in future engagements.\n
Ingestion points: The pattern_db.py script reads from append-only JSONL files located in the user's home directory (~/.claude/engagement-memory/).\n
Boundary markers: The inject command wraps the recalled intelligence in a clear Markdown header (## Prior intel) and uses structured bullet points.\n
Capability inventory: The skill allows the agent to query and update the local pattern database via the pattern_db.py script. It does not have network access or arbitrary command execution capabilities.\n
Sanitization: The skill implements comprehensive sanitization. The schemas.py script performs strict type and schema validation on every record. A looks_like_secret function uses regular expressions and entropy analysis to block common credential patterns (API keys, private keys, tokens) from being stored in the database. Additionally, the _sanitize_for_global function scrubs target-specific information before records are shared in the optional global database.\n- [DATA_EXPOSURE]: The skill manages potentially sensitive findings but employs a defense-in-depth approach to protect this data. Per-client isolation is enforced by default via the $ENGAGEMENT_DB environment variable. The secret detection guard acts as a proactive filter against accidental credential harvesting. File operations for database compaction and updates use atomic writes to prevent data corruption or partial file writes that could lead to information exposure.\n- [COMMAND_EXECUTION]: All interaction with the local database is handled through specific Python scripts using argparse for safe input handling. The skill does not utilize dangerous functions such as eval(), exec(), or os.system(), and it does not allow the injection of arbitrary shell commands.

engagement-memory