fill-model-descriptions
Pass
Audited by Gen Agent Trust Hub on Nov 17, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes model descriptions sourced from external APIs and websites. If these sources are compromised, malicious instructions could be embedded into the model metadata and influence downstream behavior.
- Ingestion points: Data is fetched from
https://openrouter.ai/api/v1/models,https://models.dev/api.json, and via web search/fetch of various model provider documentation pages. - Boundary markers: No explicit delimiters or guardrail instructions are used when interpolating external content into the
packages/llm-info/data/models.ymlfile. - Capability inventory: Across its scripts, the skill has the capability to write to the local file system and execute shell commands (
pnpm). - Sanitization: The skill applies truncation logic (first sentence, 200 character limit) which provides only minimal sanitization against adversarial sequences.
- [COMMAND_EXECUTION]: The skill executes several project-specific shell commands to synchronize models, generate code, and run tests.
- Evidence:
pnpm --filter @marimo-team/llm-info test,pnpm --dir packages/llm-info codegen, andpnpm --dir packages/llm-info sync-models --metadata-only. - [EXTERNAL_DOWNLOADS]: Fetches data from external model metadata services.
- Evidence:
GET https://openrouter.ai/api/v1/modelsandhttps://models.dev/api.json.
Audit Metadata