using-superpowers

Warn

Audited by Gen Agent Trust Hub on Apr 21, 2026

Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill uses coercive language to force compliance, including phrases such as 'ABSOLUTELY MUST', 'YOU DO NOT HAVE A CHOICE', 'This is not negotiable', and 'not optional'. These are designed to override the agent's built-in safety and logic filters regarding task appropriateness.
  • [PROMPT_INJECTION]: It explicitly commands the agent to prioritize its instructions over the default system prompt: 'Superpowers skills override default system behavior'.
  • [PROMPT_INJECTION]: The '1% rule' instructs the agent to invoke the Skill tool before any response, including clarifying questions or 'simple questions', which forces the ingestion of external data as authoritative instructions before the agent can interact with the user.
  • [PROMPT_INJECTION]: A 'Red Flags' section targets the agent's internal reasoning process, instructing it to 'STOP' and label its own logical checks (e.g., 'This is just a simple question', 'I need more context first') as 'rationalizing'.
  • [PROMPT_INJECTION]: The documentation in references/codex-tools.md recommends wrapping instructions in XML tags to ensure the model treats them as 'authoritative', which is a technique used to strengthen instruction following and potentially bypass standard safety or logic constraints.
  • [COMMAND_EXECUTION]: Mapping files in the references/ directory provide specific technical instructions for using powerful tools like shell execution (Bash/run_shell_command) and file manipulation across different platforms.
  • [PROMPT_INJECTION]: The skill creates a significant surface for indirect prompt injection: (1) Ingestion points: Mandatory Skill tool invocation reading external markdown files at the start of every task. (2) Boundary markers: No delimiters or instructions are provided to ignore malicious content within the external skills being loaded. (3) Capability inventory: The platform support described in references/ includes shell execution, subagent dispatch, and file writing. (4) Sanitization: No mechanisms for sanitizing or validating external skill content are mentioned.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 21, 2026, 09:13 AM