generate-sandbox-policy
Pass
Audited by Gen Agent Trust Hub on Nov 20, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests and processes untrusted external data (OpenAPI/Swagger specifications, markdown) to generate security policies.
- Ingestion points: Step 1 in "SKILL.md" specifies that the agent should fetch documentation from user-provided URLs or read local file paths to extract endpoint inventories.
- Boundary markers: The instructions lack specific guidance on using delimiters or ignoring natural language instructions that might be embedded within the documentation's metadata or description fields (e.g., an attacker-controlled endpoint description containing instructions to bypass restrictions).
- Capability inventory: The skill is capable of reading local files and writing/updating YAML policy files which define sensitive network and filesystem boundaries for the sandbox environment.
- Sanitization: The skill relies on "openshell policy lint" for schema validation, but this does not sanitize the semantic intent extracted from the documentation's natural language fields.
- [EXTERNAL_DOWNLOADS]: The skill fetches content from arbitrary external URLs provided by the user for the purpose of parsing API documentation. This creates a surface for the agent to interact with and process potentially untrusted or malicious web content.
- [COMMAND_EXECUTION]: The skill is instructed to execute a local CLI command ("openshell policy lint") to validate the generated policy files before they are presented or saved. This is an expected utility for policy authoring but involves executing local binaries.
Audit Metadata