redmind
Pass
Audited by Gen Agent Trust Hub on Aug 22, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill directs the agent to adopt a specialized "Red Mind" persona that overrides default safety or advisory constraints. It instructs the agent to "never look at anything neutrally" and focuses on finding ways to make systems "misbehave," which can be used to bypass safety guardrails during interactions with third-party targets.
- [COMMAND_EXECUTION]: The instructions explicitly guide the agent to perform active reconnaissance and exploitation, including fingerprinting technology, mapping attack surfaces, and executing targeted tests to achieve "demonstrable impact" such as "unauthorized shell access."
- [PROMPT_INJECTION]: The skill facilitates the ingestion of untrusted data from external targets (headers, banners, web responses) without providing safety boundaries, making it susceptible to indirect instruction injection from those targets.
- Ingestion points: External target service banners, HTTP response headers, and page content (referenced in SKILL.md and BENCHMARK.md).
- Boundary markers: Absent. The skill does not instruct the agent to use delimiters or ignore instructions embedded in the target data.
- Capability inventory: The skill encourages the use of network scanning and exploitation tools to achieve its objectives (implied in the attacker loop and benchmark).
- Sanitization: Absent. There are no instructions for filtering or validating data retrieved from targets.
Audit Metadata