Deep Scan

For the cases a per-result classifier cannot settle on its own. Run it in our infrastructure or in yours.

In the dashboard this is what Hybrid and Deep scan modes use. Hybrid calls it only where the fast scan is uncertain; Deep calls it on every response.

Where it runs

Pattern matching runs first, then a local classifier scores every result on CPU. Deep Scan reads what those two leave unsettled, which is a small enough share of the traffic that a much larger model is affordable. The stages in full.

What it adds

The classifier is deliberately conservative. It fires on payloads that resemble what it was trained on and stays silent otherwise, because something running on every tool result cannot afford to be wrong often. That leaves a real gap: attacks written to look like ordinary content, and attacks fragmented so that no single result carries a complete instruction.

Deep Scan reads for intent rather than for resemblance, so it catches cases that carry no recognisable payload at all. On Microsoft BIPIA, a benchmark of injections placed into tool-response-shaped content, it reaches an F1 of 0.918 while firing on 4 of 200 benign results (2.0%). The strongest detector we measured against it reaches 0.833, at 18.5%.

The two are complements rather than alternatives. The classifier is the cheap net that runs everywhere; Deep Scan is judgement applied to what the net cannot classify.

How we measure this

Running it

Deep Scan is a roughly 2 GB model and answers in about 150 ms on a single mid-range GPU. It is available two ways:

• Hosted by StackOne. You call an endpoint. Nothing to operate and no GPU to provision.
• In your own infrastructure. The model runs inside your boundary, so tool results never leave it. For teams where that is mandatory, not a preference.

Which one fits is usually worth a short conversation, so the form asks which way you are leaning and it is fine to say you do not know yet.

Request access

Deep Scan is in development and we are working with teams using it now. Tell us what you are protecting and we will get back to you.