Screen Browser Agent Content

Browser content is untrusted. Page text, hidden HTML, comments, and scraped snippets can include instructions aimed at the agent rather than the human-visible task.

Boundary rule

Call POST /v1/parse before page content can influence tool use, memory writes, outbound messages, or code execution.

const pageText = await browser.extractText(url);
const screen = await fetch("https://parsethis.ai/v1/parse", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.PARSE_API_KEY}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    prompt: pageText,
    metadata: {
      source_kind: "web_page",
      trust_level: "external",
      intended_action: "summarize",
    },
  }),
});

const decision = await screen.json();
if (decision.suggested_action === "block") throw new Error("Blocked page content");

Pilot check

Use the Playground Browser queue row, then compare it with the safe companion. A useful pilot should block or sandbox the hostile source while allowing the benign page-shaped companion.