SuperModel

SuperModel is WitnessAI's embeddable guardrail engine: a unified classification model that runs intent classification, jailbreak and prompt injection detection, content risk scoring, and model identity enforcement in a single inference pass, through one model and one API endpoint. Rather than protecting AI usage at the network layer, SuperModel is built to be embedded directly into another product or cloud service's own request path — one REST call in the enforcement path returns a classification result before a prompt reaches its destination LLM.

NOTE

SuperModel is a different integration path from Custom AI application defense, which screens traffic through your own WitnessAI tenant's guardrail API. SuperModel is a standalone classification engine with its own endpoint, intended for embedding guardrails into a product or platform you're building — see the SuperModel Classify API reference for the request/response contract.

The problem

Building AI guardrails from scratch into a product or platform means stitching together separate classifiers for intent, jailbreak detection, and content risk — each with its own latency, its own training pipeline, and its own blind spot. A prompt that looks benign to a standalone intent classifier can be an obvious jailbreak to a model that also sees risk signals, but independent classifiers can't share that context because each operates on its own hidden state.

How SuperModel solves the problem

SuperModel runs all four classification tasks — intent, jailbreak/prompt injection detection, risk analysis, and model identity enforcement — in one forward pass through a single model, sharing hidden-state context across all of them. That shared context lets signals correlate naturally: risk context informs intent classification, jailbreak signals shape risk scoring, and identity constraints apply across every output — at roughly 250ms of added classification latency and a 0.1% false-positive rate. The model was trained specifically for this multi-task classification role rather than adapted from a general-purpose LLM, using a combination of real-world attack patterns from enterprise deployments, synthetically generated adversarial attacks, and curated public safety benchmarks.

What capabilities SuperModel provides

Intent classification. Prompts and responses are classified against custom-defined behavioral categories using zero-shot classification — an administrator writes a natural-language definition for a category, and the model classifies against it immediately, with no retraining or labeled dataset required. Each category supports Allow, Warn, Block, or Redirect-to-a-different-model actions, and a single prompt can match multiple intents at once.
Jailbreak and prompt injection detection. Detects 10+ attack vectors — instruction-level attacks (instruction override, prefix injection, compound instructions), conversational attacks (many-shot jailbreaking, role-play manipulation, context switching), and encoding attacks (invisible/zero-width characters, emoji-based manipulation, obfuscated or non-English prompts) — using learned behavioral patterns rather than signature matching, so novel phrasings of a known attack are still caught.
Risk analysis. Scores prompts and responses, in both directions, across seven risk categories: illegal activities, harmful content, sexual content, hate speech, ethical violations (including copyright and academic dishonesty), self-harm, and fraud & deception.
Model identity enforcement. Reinforces a deployed model's assigned role and brand voice on the input side, and validates responses against those constraints on the output side — covering brand identity enforcement, competitive protection (preventing a manipulated assistant from recommending a competitor or making unauthorized commitments), and harmful response filtering.
100+ language support, including non-Latin scripts, so detection accuracy doesn't fall off outside English and a small set of Western European languages.

GuardRail Studio: configuring guardrails without code

GuardRail Studio is the companion application for building and testing the guardrail configurations SuperModel enforces. Each intent category is defined as a natural-language description (for example, "resume writing and job application assistance," or "requests to generate or debug source code") — no model retraining, dataset curation, or ML expertise required. Guardrail Studio also lets you:

Test in real time — submit sample prompts and see classification results and confidence scores live, against the same endpoint used in production.
Auto-refine guardrails — write a plain-language purpose statement and rough behavioral rules, and Guardrail Studio refines them into canonical rules with proper names and definitions, merges overlapping rules, and summarizes what changed.
Manage configurations — guardrail configs are versioned, with full CRUD available through both the UI and a REST API (the same /v1/classify endpoint the UI itself uses is open for programmatic access — see the GuardRail Studio API reference).

Deployment options

Mode	Where it runs	Who operates it
WitnessAI-hosted	WitnessAI infrastructure	WitnessAI
Partner-hosted	Partner infrastructure	Partner
Customer-hosted	Customer VPC/data center	Customer (with WitnessAI support)

In every mode, the classification hot path is self-contained — the control-plane connection used for model updates, policy sync, and audit log export runs out-of-band and does not sit in the real-time classification path, so a control-plane network break does not affect inference uptime. Customer-hosted deployments keep all AI interaction data within the customer's own boundary, which matters for regulated industries subject to cross-border data transfer restrictions (for example, GDPR Article 44, HIPAA, or regional sovereignty rules) — no AI interaction data is routed through WitnessAI.

Where SuperModel fits

The whitepaper describes three primary use cases for embedding SuperModel's classification engine:

Platform and cloud service integration — a host platform routes every AI interaction through SuperModel's classification call, applying consistent guardrails across any LLM provider or custom fine-tuned model without reconfiguring per provider.
Data-sovereign and on-premises deployments — for organizations that cannot let AI interaction data leave their own environment, SuperModel runs as a standalone deployment with the inference hot path entirely local.
Custom AI application protection — sitting between an application's orchestration layer and its LLM, SuperModel can protect customer-facing chatbots, internal copilots and coding assistants, and autonomous agents, including inspecting agent inputs before tool calls execute and reviewing outputs before an agent's action reaches the world.

How to get started

Review the API. Start with the SuperModel Classify API reference for the /v1/classify request and response contract.
Define your guardrail configuration. Write a purpose statement describing your application's intended use, and a set of rules/behaviors describing what to block — either directly in your integration's config, or authored in GuardRail Studio.
Test against representative prompts. Use Guardrail Studio's live testing, or call the classify endpoint directly with your own sample prompts, to validate classification behavior before going live.
Choose a deployment mode. Confirm with your WitnessAI representative whether a WitnessAI-hosted, partner-hosted, or customer-hosted deployment fits your data residency and operational requirements.

SuperModel The problem How SuperModel solves the problem What capabilities SuperModel provides GuardRail Studio: configuring guardrails without code Deployment options Where SuperModel fits How to get started