harness-self-improve

Pass

Audited by Gen Agent Trust Hub on Jul 8, 2026

Risk Level: SAFEDATA_EXFILTRATIONPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [DATA_EXFILTRATION]: Accesses the user's private conversation history and session transcripts in the ~/.claude/ directory. While these interaction logs are sensitive, the skill lacks network capabilities and includes privacy instructions to avoid quoting secrets or private content.\n- [PROMPT_INJECTION]: Subject to indirect prompt injection from historical data processed during the discovery phase. Malicious instructions embedded in past conversations could influence the generation of new automations. Findings: 1. Ingestion points: ~/.claude/history.jsonl and project-specific session logs. 2. Boundary markers: Absent. 3. Capability inventory: Creates files in ~/.claude/ skills/agents/commands directories and modifies settings.json. 4. Sanitization: Includes a privacy rule for secrets but lacks validation for instruction-based content.\n- [COMMAND_EXECUTION]: Instructs the agent to create automation hooks and cron jobs that modify the settings.json configuration file, enabling persistent execution of newly packaged logic.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 8, 2026, 05:03 PM
Security Audit — agent-trust-hub — harness-self-improve