AGENT LAB: SKILLS

web-search

Fail

Audited by Gen Agent Trust Hub on Feb 18, 2026

Risk Level: CRITICALREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • Remote Code Execution (CRITICAL): The documentation explicitly recommends installing the 'infsh' CLI using the command 'curl -fsSL https://cli.inference.sh | sh'. This pattern executes unverified code from a remote server, which is an untrusted source outside the defined scope of trusted organizations.
  • External Downloads (HIGH): The skill relies on external software (infsh CLI) downloaded from a non-standard source, increasing the supply chain risk.
  • Indirect Prompt Injection (LOW): The skill's primary function is to extract web content from URLs (e.g., via 'tavily/extract' or 'exa/extract') and pass it to LLMs like Claude. There are no boundary markers or sanitization steps mentioned to prevent malicious instructions embedded in web pages from hijacking the agent's behavior.
  • Evidence Chain for Category 8:
  • Ingestion points: Web content extraction tools 'tavily/extract' and 'exa/extract' in SKILL.md.
  • Boundary markers: Absent; untrusted content is interpolated directly into prompts (e.g., '').
  • Capability inventory: Network access via 'infsh' and file system access through shell redirection.
  • Sanitization: Absent; no escaping or filtering of extracted HTML/text is performed.
Recommendations
  • HIGH: Downloads and executes remote code from: https://cli.inference.sh - DO NOT USE without thorough review
  • AI detected serious security threats
Audit Metadata
Risk Level
CRITICAL
Analyzed
Feb 18, 2026, 11:22 PM