techguide

Fail

Audited by Gen Agent Trust Hub on Mar 7, 2026

Risk Level: HIGHCREDENTIALS_UNSAFEDATA_EXFILTRATIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [CREDENTIALS_UNSAFE]: The skill executes a command sequence that retrieves an active GitHub authentication token and stores it in plain text on the local filesystem.
  • Evidence: Phase 3 contains the command echo "https://GGPrompts:$(gh auth token --user GGPrompts)@github.com" > ~/.git-credentials.
  • [DATA_EXFILTRATION]: The skill writes sensitive credentials to a known sensitive file path (~/.git-credentials), which constitutes unauthorized exposure of authentication secrets in a persistent manner.
  • Evidence: The use of git config --global credential.helper store combined with the direct writing of the token to the home directory ensures the credential persists beyond the session.
  • [COMMAND_EXECUTION]: The skill performs shell commands to modify global system configurations and execute network operations.
  • Evidence: Use of git config and git push commands in Phase 3.
  • [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it processes and incorporates untrusted external data into the agent's context and final output.
  • Ingestion points: The skill reads styles/{name}.html (where {name} is a user-supplied argument) and aggregates content from WebSearch and context7 MCP documentation queries in Phase 1.
  • Boundary markers: Absent. There are no instructions or delimiters used to prevent the agent from executing commands that might be embedded in the researched documentation or style files.
  • Capability inventory: The skill has the ability to modify local files via the Edit tool and push changes to remote repositories via git push.
  • Sanitization: Absent. Content retrieved from the web or documentation tools is directly passed to subagents to be formatted into the final HTML guide.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Mar 7, 2026, 05:58 AM