Protect an AI agent from prompt injection in fetched web pages

How can I scan a web page or text for prompt injection before my AI agent reads it?

Send the text, or the page URL, to POST /v1/safety/injection-scan before the agent reads it. It returns a risk level (none, low, medium or high) and each finding with its kind, offset and an excerpt: hidden instructions, chat-template and role markers, tool-call bait, requests for keys or seed phrases, exfiltration links and invisible characters.

Example: a comment with a hidden instruction

Example call: safety-injection-scan

Body

Injection scan body
FieldValues
textText to scan, up to 100,000 characters
urlA page to fetch and scan, hidden parts included. Send this or text.

What it detects

  • Hidden instructions such as “ignore previous instructions”, role impersonation and chat-template markers.
  • Tool-call bait, requests for API keys or seed phrases, and links that would exfiltrate data.
  • Invisible text: zero-width characters, bidi overrides, Unicode tag characters, variation-selector payloads, hidden HTML elements and HTML comments, long base64 blobs.

The scan is pattern-based with no model: fast and deterministic, but risk none does not prove text is safe. It lowers the risk; it does not make third-party text safe to obey.

Scanning built into page reading

GET /v1/web/read runs the same kind of check on every page it reads and reports it in injection_risk and injection_warnings, without removing anything. Every paid call through the Octio gateway is also scanned before it is charged: a provider response that looks like instructions aimed at agents fails with UPSTREAM_SUSPICIOUS_CONTENT and costs nothing.

FAQ

What is prompt injection?

Text written to make a language model follow new instructions, hidden in content the model reads, such as a web page, a PDF, an email or a tool result.

Does the scanner use an LLM?

No. It uses patterns and character checks, so the same input always gives the same answer, quickly.

Can it scan a URL directly?

Yes. Send url instead of text and it fetches the page and scans it, hidden parts included.

Updated 2026-11-16.