Is This a Safe WebsiteAn agent reads untrusted instructions too
External pages can contain instructions aimed at changing an agent’s behaviour. Treat retrieved content as data with an untrusted origin. URL reputation alone cannot prove the absence of prompt injection or authorise an agent to disclose secrets or perform actions.
Browsing policy example
Permit an agent to read only approved public sources for the named task. Do not grant shell, email, payment or credential access merely because browsing is allowed. Require human approval for material actions. Validate each redirect destination at the actual browsing gateway; this local URL checker does not follow redirects.
Evidence and operational review
Log allowed destinations and action decisions through an approved privacy-aware process. Test a fictional page that asks the agent to send a secret to another address: the expected outcome is no disclosure and a reported untrusted instruction. Record the task, source, enforced permission boundary and observed outcome.