
Argent with an ONNX injection classifier
Argent scores prompts for injection and jailbreak attempts with a small classifier model that runs in ONNX inside your own deployment. It runs on plain CPUs, the weights are fetched and pinned by you, and no text is sent anywhere to be judged.
What Argent does with the classifier
A prompt-injection check that runs on the machines you already have.
OWASP ranks prompt injection first, as LLM01, in its Top 10 for LLM Applications 2025 (OWASP GenAI Security Project). The UK National Cyber Security Centre warns that prompt injection is not SQL injection and may be worse (NCSC, December 2025).
ONNX is an open format for machine-learning models. Argent serves injection models in it from a sidecar next to the gateway.
- One pass per text. The model gives one injection probability, which Argent applies to both the prompt_injection and jailbreak categories.
- Scores what a person wrote. Argent pulls out the prompt text before scoring, so the JSON around it is not mistaken for an attack.
- Long prompts in windows. A long prompt is scored in overlapping windows, and the verdict is the highest window. Text past the window cap is marked unscored, never assumed safe.
- Blocks or watches. Above your threshold, a prompt is refused before it reaches the model, or, in observe mode, scored and recorded without blocking.
The models it serves
One sidecar can serve several models. Each has reviewed pins and a known license.
| Model | License | Runs on |
|---|---|---|
| ProtectAI deberta-v3-base-prompt-injection-v2 | Apache-2.0 | CPU. The default injection model on GPU installs |
| PIGuard | MIT | CPU |
| Horizon prompt-injection-guard-base | Apache-2.0 | CPU |
- Pinned to a revision. Every model is fetched at a pinned revision and every file is checked against its SHA-256. The sidecar refuses a model whose revision does not match.
- No remote code at run time. Where a model needs export code, it runs once on the operator's machine during export, never inside the sidecar.
- Checked export. Models converted to a smaller format are compared with the original, and an export whose verdicts change is refused.
How it connects
The classifier is a service inside your Argent deployment. Nothing leaves it.
| Where it runs | A sidecar service beside the gateway, onnx-classifier on port 8080 |
| Calls | The gateway posts text to /score; /models lists each loaded model with its revision and hash |
| Docker Compose | The classifier-onnx profile, or classifier-onnx-gpu on a GPU machine |
| Helm | classifier.onnx.enabled=true; a fetch Job pulls the models |
| Weights | Fetched by you and never baked into any image |
What you configure
Fetch the weights, switch the sidecar on, pick the model, then set the policy.
- Accept the license and fetch. Run
make accept-onnx-license, thenmake fetch-onnx-model. In Helm, the fetch Job runs only onceclassifier.onnx.fetch.licenseAccepted=trueis set. - Switch the sidecar on. Use the Compose profile or the Helm value above.
- Pick the model. A CPU-only install ships with no default classifier, and the console says so plainly until an admin picks one. A GPU install defaults to the ProtectAI model for injection.
- Set the tenant's classifier policy. On the Classifier page: which categories, the block threshold, block or observe, what happens on a timeout, and on or off.
- Watch first. In watch, would-be blocks are recorded and nothing is refused. Switch to enforce when the record looks right.
Air-gapped? This is the recommended classifier for CPU-only and air-gapped installs: no GPU, no outside call. See Air-Gapped Install.
Honest about what it scored
A prompt that was not scored is recorded as not scored. It is never recorded as clean.
- No model, no false comfort. Tenants with no model picked are recorded as not scored, never as clean, and are never blocked for it.
- Unknown model fails closed. A request for a model the sidecar does not have is treated as could not measure.
- Every change is announced. When a deployment default takes effect for a tenant, Argent signs one event naming the model, and the tenant keeps its own watch or enforce setting.
Where this fits
Pages that go deeper.
- Run a guardrail already? Hand the decision to LLM Guard or Amazon Bedrock Guardrails instead.
- Capabilities: Prompt Injection Defense, Response Screening, Watch Before You Enforce, Session Replay and all capabilities.
- Goals: threat defense, data loss prevention, sovereign AI and all goals.
- Use cases: governing customer-facing AI assistants, rolling out copilots and chat safely, securing AI coding agents, controlling agent tool access and AI in regulated and air-gapped environments.
- Where it runs: air-gapped on-premises, private cloud or your public cloud VPC. Compare them on deployment options.
- Industries: Financial Services, Healthcare, Insurance, Public Sector, Retail and all industries.
- More integrations: every system Argent connects to, grouped by type, on the integrations page.
Want this running on your own use cases?
All downloads