agent-browser-security-production-safety
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill identifies and attempts to mitigate the risk of indirect prompt injection by processing untrusted browser data.
- Ingestion points: Browser page text, console output, network payloads, and downloaded files (SKILL.md, 01-auth-state-secrets-trust.md).
- Boundary markers: Explicit instructions are present to treat all browser content as untrusted and to disregard page-provided instructions that conflict with the safety contract.
- Capability inventory: Uses the
agent-browserCLI to interact with the environment and process files. - Sanitization: Strictly requires the redaction and masking of PII, secrets, and authentication tokens before sharing artifacts.
- [COMMAND_EXECUTION]: The skill incorporates the use of the
agent-browserCLI tool, instructing the agent to execute commands for version checking and loading local skill guidance (SKILL.md, Step 2).
Audit Metadata