Fetch Guard: strip prompt injection from headers and pages before your agent reads them

The basic safe-fetch MCP server already exists several times over. There is a project actually called Fetch Guard, plus Safer Fetch, webfetch-mcp and FetchSandbox, and Lasso and Snyk/Invariant sell response-scanning MCP gateways to enterprises. So a new entrant can't win on stripping hidden DOM text and invisible Unicode, and the Fetch Guard name is already in use. The clearest gap is HTTP response headers. Recent measurement work found roughly half of in-the-wild injection instances in headers, which DOM-focused tools miss. None of the fetch servers I found clearly advertises header allowlisting or header-specific detection. A rules feed built on crawl data of real header and meta-tag injections, updated continuously and shipped as a feed any fetch tool or gateway can consume, would be more defensible than one more open-source MCP server. That feed is the part agent builders might pay for. The test-pages idea overlaps with Promptfoo's indirect-web-pwn, Perplexity's BrowseSafe-Bench and IPI-proxy. However, these are generators or datasets rather than a public, stable set of live URLs with planted header injections. A hosted conformance site would be cheap to build and would market the feed well, but it is unlikely to be a business on its own. Two risks: indie developers rarely pay, and the same study found models seldom obey header injections. That weakens the urgency unless you can show real exploits.

• Fetch Guard (Glama listing)

  An MCP server and CLI with the same name as your idea. It returns markdown ready for an LLM. Its own description says it strips hidden text, invisible Unicode and off-screen elements, and flags prompt-injection attempts. Both the name and the core feature are already taken. I found nothing independent that verifies how well it works.

• Safer Fetch MCP Server

  A fork of Anthropic's reference fetch server that adds Goose's injection-safeguard patterns. It wraps fetched content in boundary tags with a random ID and scans for suspicious patterns.

• webfetch-mcp

  A fetch MCP server that can sanitize prompt injection in two modes: flag it or strip it.

• FetchSandbox MCP

  Sold as secure web access for AI agents through MCP. It is another drop-in fetch replacement with a security angle.

• Lasso MCP Gateway

  An enterprise MCP gateway that filters requests and responses on every tool call, including scanning for injected instructions. It works across all tools, not just fetch.

• Invariant MCP-Scan proxy (Snyk)

  Its proxy mode applies guardrail policies and custom rules to tool calls and responses. Snyk acquired Invariant in 2025.

• Perplexity BrowseSafe and BrowseSafe-Bench

  An MIT-licensed detector model plus a benchmark of about 14.7k samples. The benchmark plants injections in HTML comments, data attributes, hidden text and form fields. It overlaps with the test-pages idea.

• Promptfoo indirect-web-pwn

  A red-team harness that generates web pages with hidden payloads (CSS-hidden or invisible text) and checks whether an agent follows them. It is a direct competitor to the injection test pages.

• IPI-proxy

  A research intercepting proxy that injects attack payloads into live pages to red-team browsing agents. It tests defenses rather than providing one.

• Header injection measurement study (trackr.live write-up)

  Not a product. It reports that about 51% of the 15,387 confirmed injection instances it found arrived in HTTP response headers, which DOM-only scanners miss. That supports the header angle, but the same study found models rarely obeyed these injections.

Checked 2026-10-16