agent-browser

Warn

Audited by Gen Agent Trust Hub on Oct 15, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill instructs the agent to install the agent-browser package globally using npm i -g agent-browser, which introduces an external dependency from a public registry.
  • [COMMAND_EXECUTION]: The skill makes extensive use of the agent-browser CLI to perform browser automation tasks, including opening URLs, managing sessions, and capturing screenshots.
  • [DYNAMIC_EXECUTION]: The skill uses agent-browser skills get core --full to retrieve and follow instructions dynamically from the installed CLI tool at runtime. This allows the CLI to provide arbitrary instructions to the agent that are not present in the skill's static instructions.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to interact with and scrape web pages, which serves as an ingestion point for untrusted external content.
  • Ingestion points: Browser accessibility tree snapshots obtained through agent-browser snapshot -i in SKILL.md.
  • Boundary markers: None identified to separate scraped web content from agent instructions.
  • Capability inventory: Execution of shell commands via agent-browser and file system writes for media evidence.
  • Sanitization: No mention of sanitization or filtering of ingested web data before processing.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 15, 2026, 11:43 AM