generate-sandbox-policy
Pass
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external, potentially untrusted data to generate security-critical configurations.
- Ingestion points: Reads local project files (e.g.,
sandbox-policy.yaml) and fetches external API documentation (OpenAPI/Swagger specs, markdown) from user-provided URLs or web search results. - Boundary markers: The instructions do not specify the use of delimiters or 'ignore embedded instructions' markers when processing ingested API documentation.
- Capability inventory: The skill is capable of performing file system writes to update security policies and generating L4/L7 network rules.
- Sanitization: There is no mention of explicit sanitization or validation of the external documentation content before it is used to formulate policy rules.
- [EXTERNAL_DOWNLOADS]: The skill fetches documentation and schema references from remote sources.
- Evidence: Fetches configuration and endpoint schemas from the vendor's official documentation at
docs.nvidia.comand retrieves documentation from well-known services such as GitHub, Anthropic, and OpenAI to automate policy creation. - [DATA_EXFILTRATION]: The skill reads local sandbox policy files which contain sensitive network and process configuration details.
- Evidence: The workflow involves reading existing policy files (e.g.,
sandbox-policy.yaml) to identify current state and merge changes, exposing existing security controls to the agent's context.
Audit Metadata