sglang
Pass
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user queries and articles to drive structured generation and agent workflows.
- Ingestion points: External data enters the agent context through the
user_query,article, andtextparameters in the SGLang function examples found inSKILL.mdandreferences/structured-generation.md. - Boundary markers: The prompt templates use natural language labels (e.g.,
User:,Assistant:,Article:) but do not include explicit security instructions or delimiters to prevent embedded instruction overrides. - Capability inventory: The skill's logic is constrained to LLM inference and structured output via the SGLang library; no dangerous capabilities like arbitrary filesystem writes or unauthorized network requests are exposed.
- Sanitization: External content is interpolated directly into prompts without specific filtering or escaping mechanisms.
- [EXTERNAL_DOWNLOADS]: The documentation provides instructions to install dependencies from the official SGLang GitHub repository and the FlashInfer project's specialized wheel registry.
Audit Metadata