web-research-workflow

Warn

Audited by Gen Agent Trust Hub on Mar 30, 2026

Risk Level: MEDIUMCREDENTIALS_UNSAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [CREDENTIALS_UNSAFE]: The 'rules/browser-patterns.md' file recommends saving browser session states to '/tmp/session-example.json'. Because the /tmp directory is generally world-readable on multi-user systems, session cookies and tokens stored there could be accessed by unauthorized users.\n- [DATA_EXFILTRATION]: The skill transmits search queries and URL data to 'api.tavily.com' to utilize its semantic search and content extraction capabilities. This is a well-known service for AI research.\n- [COMMAND_EXECUTION]: The workflow makes use of 'agent-browser eval' to run JavaScript within the context of target web pages for data extraction purposes. While intended for scraping, this provides a mechanism for script execution in a browser environment.\n- [PROMPT_INJECTION]: By design, this skill processes untrusted content from the web, creating an attack surface for indirect prompt injection. Malicious instructions on a target website could attempt to influence the agent's behavior or exfiltrate data.\n
  • Ingestion points: External web content retrieved via 'WebFetch', 'Tavily', and 'agent-browser'.\n
  • Boundary markers: The instructions do not include specific delimiters or warnings to isolate untrusted content from the agent's primary reasoning logic.\n
  • Capability inventory: The skill utilizes tools such as 'Bash', 'Read', 'Write', and 'WebFetch', which could be leveraged if an injection occurs.\n
  • Sanitization: The skill lacks explicit sanitization or validation steps for content fetched from remote servers before it is processed.\n- [EXTERNAL_DOWNLOADS]: The skill is configured to fetch data from arbitrary external URLs as part of its web research and competitor monitoring workflows.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 30, 2026, 03:24 AM