social-coach

Pass

Audited by Gen Agent Trust Hub on Aug 24, 2026

Risk Level: SAFE
Full Analysis
  • [PROMPT_INJECTION]: The skill explicitly prohibits PUA and manipulative tactics, enforcing a safety bottom line for the agent's behavior. No adversarial prompt patterns or attempts to bypass safety filters were detected.
  • [DATA_EXFILTRATION]: Social interaction data is stored locally in the ~/.openclaw/workspace/memory/social-coach/ directory. The skill encourages the use of aliases for individuals rather than real names. There are no network requests or patterns indicating data exfiltration to external servers.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied interaction data stored in local JSONL files. Ingestion points: User input for social records via commands like /记录邀约 or /破冰. Boundary markers: Data is stored in JSONL format, but there are no specific instructions to ignore embedded commands within the text fields. Capability inventory: The agent utilizes read and write tools for file system interaction within its workspace. Sanitization: Absent.
  • [COMMAND_EXECUTION]: The skill uses file reading and appending tools to manage its local memory. These operations are scoped to a specific workspace directory and represent standard functional behavior for an agent with persistent memory.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 24, 2026, 04:40 PM
Security Audit — agent-trust-hub — social-coach