access-review
Pass
Audited by Gen Agent Trust Hub on Jul 15, 2026
Risk Level: SAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external identity and entitlement data (e.g., role names, group descriptions) which constitutes an untrusted input surface. However, the skill includes extensive 'Injection Hardening' and 'Prompt Injection Safety' sections that instruct the agent to treat all data as untrusted, avoid executing instructions found in metadata, and flag attempts to override instructions as security findings.
- [COMMAND_EXECUTION]: The skill specifies 'Read', 'Grep', and 'Glob' as allowed tools. These are used appropriately for searching and analyzing configuration files. There are no patterns suggesting arbitrary or dangerous command execution.
- [DATA_EXFILTRATION]: The instructions explicitly prohibit the exfiltration of user lists, credentials, or entitlement data discovered during reviews. The external references provided point to legitimate, well-known government (NIST) and security (CIS, ISACA) documentation sites.
- [PROMPT_INJECTION]: Static analysis flagged the presence of instructions related to overriding guidelines. However, context reveals these are defensive instructions teaching the agent to detect and ignore such patterns in the data it audits, rather than attempts to bypass the agent's own safety filters.
Audit Metadata