skills/raychang/agent-skills/cove/Gen Agent Trust Hub

cove

Pass

Audited by Gen Agent Trust Hub on May 6, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill processes untrusted user input or previous conversation history and passes it to secondary agent processes.\n
  • Ingestion points: User-supplied text provided via the /cove command (Step 1) and the most recent assistant response in the conversation history.\n
  • Boundary markers: The prompt template for subagents in Step 3 (Question: <the verification question>) lacks clear structural delimiters (such as XML tags or markdown blocks) or 'ignore instructions' wrappers to prevent the verification question from containing instructions that could manipulate the subagent's logic.\n
  • Capability inventory: The skill instructs the agent to utilize invoke_agent (Gemini CLI) or the Agent tool (Claude Code) to execute sub-processes with general-purpose capabilities.\n
  • Sanitization: There is no evidence of validation, escaping, or filtering of the claims extracted from the draft before they are interpolated into the prompt sent to the subagent.
Audit Metadata
Risk Level
SAFE
Analyzed
May 6, 2026, 06:33 AM
Security Audit — agent-trust-hub — cove