The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches dependencies and source code from external repositories and specialized indices.
Evidence: Installation commands include git clone https://github.com/sgl-project/sglang.git and pip install sglang[all] flashinfer -i https://flashinfer.ai/whl/cu121/torch2.4/ in SKILL.md and references/deployment.md.
[INDIRECT_PROMPT_INJECTION]: The skill provides numerous patterns for ingesting untrusted data into LLM context, which represents an attack surface for indirect prompt injection.
Ingestion points: Untrusted content is interpolated into prompts via variables like user_query, text, article, description, and natural_language across SKILL.md and references/structured-generation.md.
Boundary markers: The examples utilize basic text prefixes such as 'User:', 'Assistant:', or 'Article:' which are insufficient to prevent sophisticated injection attacks.
Capability inventory: The skill enables LLM generation, tool call execution (e.g., in the agent_workflow), and structured output generation (JSON, Regex, EBNF Grammars).
Sanitization: No sanitization, escaping, or schema validation for the raw input data is demonstrated before it is interpolated into the prompts.

sglang