ai-security-analyst
Pass
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONNO_CODE
Full Analysis
- [PROMPT_INJECTION]: The documentation includes common injection patterns such as 'Ignore previous instructions' and 'DAN' (Do Anything Now) mode. These are presented strictly as examples within the 'Prompt Injection Assessment' taxonomy (Phase 2) for the purpose of testing other AI deployments, rather than as instructions for the agent to follow during its own operation.
- [INDIRECT_PROMPT_INJECTION]: The skill identifies ingestion surfaces where an agent might process untrusted data, such as PDFs, web pages, and database records.
- Ingestion points: External documents, browser tools, and database queries during the security review process (Phase 2).
- Boundary markers: The skill explicitly recommends defense validation including 'Instruction hierarchy' and 'Scope enforcement' to prevent context leakage.
- Capability inventory: No subprocesses, file writes, or network operations are defined in the skill's source files.
- Sanitization: Recommends pre-processing controls like PII masking and redaction to protect sensitive data.
- [NO_CODE]: The skill is composed entirely of instructional Markdown text. It does not contain any Python scripts, Node.js code, or references to external executable dependencies, significantly reducing the potential for runtime exploitation.
Audit Metadata