Safe web browsing tools for AI agents

How can an AI agent browse the web without falling for phishing or prompt injection?

Put three checks between the agent and the web: check each link with GET /v1/safety/url-check before opening it, read the page as clean text with GET /v1/web/read (or /v1/docs/pdf-text for PDFs), and scan anything else the agent ingests with POST /v1/safety/injection-scan. Then pass the content to the model as data, never as instructions.

Example: checking a link first

Example call: safety-url-check

Example: reading the page

Example call: web-read

What these checks do not do

  • A clean URL check is not proof a page is safe: new phishing pages appear before any feed lists them.
  • The injection checks are pattern-based. Risk none does not prove text is safe to follow.
  • None of them replaces narrow tool permissions and a person’s approval for actions that move money or data.

FAQ

Why read pages through an API instead of a browser?

The agent gets only the main text as Markdown, with no scripts running and with injection warnings attached, and failed reads are not charged.

Which threat feeds are used?

URLhaus by abuse.ch and PhishTank, downloaded in the background, plus static signals read from the URL.

Updated 2026-11-16.