Runs a curated signature corpus over a piece of untrusted text — content
an agent is about to consume, a retrieved document, a tool result, an
email body — and returns the matched injection patterns plus a bounded
0..1 risk score. This is a signal, never a policy decision: the caller
decides what to do with a flagged input.
Detected classes: instruction_override (ignore/override previous rules),
role_reassignment (you are now DAN / developer mode), exfiltration (leak
the system prompt or a secret to a URL), tool_smuggling (covertly invoke
a tool, delete/destroy data), boundary_spoof (fake system/assistant turn
delimiters). Input is normalized first to blunt cheap evasions (zero-width
characters, smart quotes, whitespace padding).
Use this tool when:
- You are an agent about to feed retrieved or third-party text into a model
and want to check it for embedded instructions first.
- You are triaging why a tool description or web page looks suspicious.
Do NOT use this tool when:
- You want a trust verdict on a domain or entity — use `cross_lens_verify`.
- You want to scan a whole MCP server's tools — use `scan_mcp`.
Inputs:
- `text` (body, required): the untrusted text to scan. Max 200,000 chars.
Returns:
- `flagged`: true if any signature matched.
- `score`: bounded 0..1 risk score (saturating — one high-severity hit is
already strongly flagged; many hits approach but never exceed 1).
- `severity_max`: highest severity among matches (`high`/`medium`/`low`) or null.
- `classes`: distinct injection classes matched.
- `matches`: each matched signature `{ id, class, severity, excerpt }`.
Cost:
- Free. No API key required. Pure edge computation, no external calls.
Latency:
- Typical <20ms.
Connector