web-search-tavily
Pass
Audited by Gen Agent Trust Hub on Jun 23, 2026
Risk Level: SAFEDATA_EXFILTRATIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill instructions direct the agent to read sensitive information (API keys) from the local filesystem at
~/.gooseworks/credentials.json. While this is intended for authenticating with the vendor's proxy service, it establishes a pattern of accessing local secrets. - [COMMAND_EXECUTION]: Setup steps involve the execution of shell commands using
python3 -cto parse local configuration files and export environment variables (GOOSEWORKS_API_KEYandGOOSEWORKS_API_BASE). - [EXTERNAL_DOWNLOADS]: The documentation references
npx gooseworks loginandnpx goose-skills install, which use the Node Package Executor to download and run code from external registries. - [PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by retrieving and processing untrusted data from the web through search, extraction, and crawling capabilities.
- Ingestion points: Web search results, extracted page content (markdown/text), and site-wide crawl data retrieved via the Tavily API.
- Boundary markers: Absent. The instructions do not define delimiters or provide warnings to the agent to disregard instructions embedded within the retrieved web data.
- Capability inventory: The skill allows the agent to perform outbound network requests using
curlto interact with the Tavily proxy and ingest the results. - Sanitization: No sanitization, filtering, or validation of the remote content is specified before it is interpolated into the agent's context.
Audit Metadata