speculative-decoding
Pass
Audited by Gen Agent Trust Hub on Sep 22, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill instructs the user to download and install code from external research repositories.
- Fetches the Medusa inference acceleration framework from the 'FasterDecoding' organization on GitHub.
- Fetches the Lookahead Decoding implementation from the 'hao-ai-lab' organization on GitHub.
- [INDIRECT_PROMPT_INJECTION]: The skill processes external text prompts through inference routines, creating a potential surface for indirect prompt injection.
- Ingestion points: Prompts are ingested through variables in the Python code snippets defined in 'SKILL.md'.
- Boundary markers: The provided implementation examples lack explicit boundary markers or delimiters to isolate user input from the rest of the generation process.
- Capability inventory: The skill invokes model generation capabilities, including 'medusa_generate' and 'lookahead.generate', to process the input data.
- Sanitization: There is no evidence of input validation, filtering, or sanitization logic in the provided code snippets.
Audit Metadata