op-injection-scanner

An MCP server for prompt injection boundary enforcement. AI agents call scan_url(url) instead of fetching URLs directly. The server fetches the URL (with SSRF protection), strips HTML to plain text, and scans the content for prompt injection using a three-tier model escalation strategy (Haiku → Sonnet → Opus). Results are cached by URL and TTL bucket so repeated calls within the cache window incur no additional LLM cost.

Installation

git clone https://github.com/ethereum-optimism/op-injection-scanner
cd op-injection-scanner
bun install

Claude Code MCP config

Add to ~/.claude.json or your project's .claude/settings.json:

{
  "mcpServers": {
    "op-injection-scanner": {
      "command": "bun",
      "args": ["run", "/absolute/path/to/op-injection-scanner/src/server.ts"]
    }
  }
}

The server requires ANTHROPIC_API_KEY in the environment.

Usage

Agents call scan_url with a URL instead of fetching it directly:

scan_url({ url: "https://some-external-docs.example.com/page" })

Clean response — status 200, tool returns:

{
  "status": "clean",
  "content": "The stripped page text...",
  "url": "https://some-external-docs.example.com/page",
  "scanned_at": "2026-03-08T12:00:00.000Z",
  "model_used": "claude-haiku-4-5",
  "escalated": false
}

Blocked response — tool returns isError: true so the calling agent receives a tool error:

{
  "status": "blocked",
  "reason": "Content contains instructions attempting to override agent behaviour",
  "url": "https://malicious.example.com/page",
  "scanned_at": "2026-03-08T12:00:00.000Z",
  "model_used": "claude-haiku-4-5",
  "escalated": false
}

When the tool returns isError: true, the calling agent should not process the URL further.

Environment variables

Variable	Default	Purpose
ANTHROPIC_API_KEY	(required)	Anthropic API key — read automatically by the SDK
CACHE_MAX_ENTRIES	500	Maximum number of cached scan results
CACHE_TTL_SECONDS	3600	Cache entry lifetime in seconds (1 hour)
CONFIDENCE_THRESHOLD	0.8	Minimum confidence for a verdict — below this, escalate to next tier
FETCH_TIMEOUT_MS	10000	HTTP request timeout in milliseconds
MAX_RESPONSE_BYTES	2097152	Maximum response body size (2 MB)
MAX_REDIRECTS	5	Maximum number of HTTP redirects to follow
MAX_URL_LENGTH	2048	Maximum URL length in characters

Design decisions

Why MCP, not a hook

A Claude Code hook fires automatically on every tool call, but it cannot be selectively invoked or tested in isolation. An MCP server is explicit — agents opt in to scanning, the interface is testable, and it can be extended (new tools, new scan types) without touching hook infrastructure.

v1 scope

This version detects prompt injection only. Malicious intent detection (e.g. phishing, credential harvesting) is a future iteration. The clean verdict means "no injection attempt detected" — not "content is safe in all respects."

Tiered model strategy

• Haiku — fast, cheap, handles the majority of clear-cut cases
• Sonnet — escalated when Haiku returns uncertain or confidence < 0.8
• Opus — final arbiter, called only when Sonnet also cannot reach a confident verdict; schema enforces a definitive clean or blocked verdict (no uncertain allowed); parse failure defaults to blocked

No auth or rate limiting

This is internal dev tooling. Network isolation is the access control. Auth and per-session rate limiting are noted as future work in the source.

DNS rebinding (known v1 limitation)

The SSRF guard resolves DNS once before the request and checks all returned IPs. A sophisticated attacker controlling a low-TTL DNS record could return a public IP on the initial check and re-resolve to a private IP at TCP connection time, bypassing the guard. Full mitigation requires a custom DNS-pinning HTTP client. This is acceptable for internal tooling but must be addressed before any public or shared deployment.