site-teardown
Pass
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill fetches and processes untrusted external content (HTML, JavaScript, and CSS) from arbitrary URLs provided by the user or extracted from the source code. This data is interpolated into prompts (e.g., dynamic lists of class names) and fed back into the agent's context to generate a teardown report.
- Ingestion points: The skill uses
WebFetchto retrieve the initial URL, main JavaScript files, and main CSS files. - Boundary markers: While the skill provides specific extraction instructions to the model, it lacks explicit safety delimiters or "ignore embedded instructions" warnings for the untrusted content it ingests.
- Capability inventory: The skill has the capability to read from the network (
WebFetch) and write to the local file system (research/directory). - Sanitization: There is no evidence of sanitization, escaping, or validation of the external content before it is processed by the LLM.
- [EXTERNAL_DOWNLOADS]: The skill is built around downloading resources from external, unverified websites. While this is the intended purpose for site analysis, fetching arbitrary JavaScript and CSS files from the web introduces risks if those files are malicious or contain exploit code targeting the agent's processing logic.
Audit Metadata