taste-review

Warn

Audited by Gen Agent Trust Hub on Aug 20, 2026

Risk Level: MEDIUMCREDENTIALS_UNSAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [CREDENTIALS_UNSAFE]: The script uses the macOS security find-generic-password command to access the system Keychain. It specifically targets an entry labeled "Claude Code skill OAuth" to retrieve a password/token. While used for legitimate authentication of the CLI tool, programmatic access to system credential stores is a sensitive data exposure concern.
  • [COMMAND_EXECUTION]: The skill executes the claude CLI via a bash script (claude -p "$prompt"). This behavior involves spawning subprocesses to handle user-supplied questions and repository file content.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data (user questions and file content) and passes it to an external LLM call via the CLI.
  • Ingestion points: The shell script template ingests a <your question> string and <paths> to files within the repository.
  • Boundary markers: None; the input is interpolated directly into an EOF-delimited string block without specific escaping or ignore-instruction delimiters.
  • Capability inventory: The skill can execute the claude CLI, access the macOS Keychain, and read local files.
  • Sanitization: None; the content from local files and the user question are passed directly to the claude command line as a prompt argument.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Aug 20, 2026, 10:22 PM
Security Audit — agent-trust-hub — taste-review