proactive-agent

Pass

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: LOW
Full Analysis
  • [PROMPT_INJECTION] (SAFE): The skill implements multi-layered defensive instructions across assets/AGENTS.md, assets/SOUL.md, and references/security-patterns.md, explicitly training the agent to treat external content as passive data rather than executable instructions.\n- [COMMAND_EXECUTION] (LOW): The scripts/security-audit.sh utility executes local shell commands (stat, grep) to verify environment security. These operations are restricted to static patterns and do not incorporate untrusted external input into command strings.\n- [DATA_EXFILTRATION] (SAFE): Operational rules strictly prohibit sending data externally without explicit human approval. scripts/security-audit.sh specifically checks for credential exposure and loops back gateway configurations to prevent unintended exposure.\n- [INDIRECT_PROMPT_INJECTION] (LOW):\n
  • Ingestion points: Processes external data from web searches, emails, and PDFs as defined in assets/AGENTS.md.\n
  • Boundary markers: Employs clear instructional boundaries and 'Heartbeat' checks in assets/HEARTBEAT.md to scan for injection patterns in recently processed data.\n
  • Capability inventory: Includes file system exploration, web searching, and execution of the included audit shell script.\n
  • Sanitization: Relies on behavioral monitoring and instruction-level isolation rather than programmatic string sanitization.
Audit Metadata
Risk Level
LOW
Analyzed
Feb 17, 2026, 04:22 AM