guidewire-security-and-rbac

Warn

Audited by Gen Agent Trust Hub on Sep 9, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The skill utilizes a dynamic shell execution pattern where output from a decryption tool (SOPS) is transformed via sed and then executed using eval to load environment variables. Additionally, the skill describes processes that automatically generate shell scripts (e.g., scripts/sops-env) to the local file system.
  • [INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of external, untrusted data from the Guidewire Cloud API (including claim narratives and contact information) which can influence agent behavior if not properly delimited.
  • Ingestion points: Data is read from multiple Guidewire Cloud API resources such as Claims, Policies, and Contacts as described in SKILL.md and API_REFERENCE.md.
  • Boundary markers: The instructions do not provide explicit delimiters or instructions for the agent to treat data from these external resources as untrusted during processing.
  • Capability inventory: The skill is granted Bash(curl:*) for network communication and Read/Write/Edit tools for file system interaction.
  • Sanitization: The skill includes a redactPii Typescript implementation intended to sanitize sensitive fields (SSN, DOB, etc.) before they are written to logs.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 9, 2026, 03:46 AM
Security Audit — agent-trust-hub — guidewire-security-and-rbac