token-coach

Fail

Audited by Gen Agent Trust Hub on Oct 11, 2026

Risk Level: HIGHPERSISTENCEDYNAMIC_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PERSISTENCE]: The skill implements a 'Keep-Warm' feature that modifies system configuration by installing a scheduler on macOS to maintain active prompt caches via background pings. This ensures the skill's logic continues to run outside of the active agent session.\n- [DYNAMIC_EXECUTION]: The skill dynamically resolves the path to an executable Python script ('measure.py') by searching through multiple directories, including shared plugin caches like '$HOME/.claude/plugins/cache'. It then executes this script using 'python3' with various subcommands. This pattern is risky as it could execute malicious code if an attacker manages to place a file in the searched cache locations.\n- [COMMAND_EXECUTION]: Core functionality relies on executing complex shell commands, including 'find -L', 'sed', and 'python3' calls. These are used to discover system paths and collect diagnostic data from the environment.\n- [PROMPT_INJECTION]: The skill's instructions contain a specific behavioral override that commands the agent to 'NEVER recommend cutting our own tools' and to exclude the tool's own components from any optimization or consolidation suggestions, regardless of their actual usage or cost. This bypasses the tool's primary purpose of providing objective token optimization advice.\n- [INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect injection as it processes external data to drive its coaching logic.\n
  • Ingestion points: Data is ingested from the JSON output of 'measure.py' subcommands, local environment variables, and project files in the current working directory.\n
  • Boundary markers: The instructions do not define clear delimiters or escaping for the data parsed from script outputs or local files.\n
  • Capability inventory: The skill has the capability to execute shell commands, read and move files, and install system schedulers.\n
  • Sanitization: There is no evidence of sanitization or validation performed on the diagnostic data before it is presented to the agent for decision-making.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Oct 11, 2026, 09:38 PM