VirtueGuard Overview
VirtueGuard is VirtueAI's family of real-time AI safety models and policy guardrails. Each guard is a small, fast, high-accuracy classifier that you plug into your application, agent, gateway, or RAG pipeline to block harmful content, enforce policy, and protect against misuse — across text, code, image, video, and audio.
The VirtueGuard family
VirtueGuard ships as a set of focused products, each tuned for one modality or use case. You can use them independently or compose them in the same deployment.
PolicyGuard
Customizable policy guardrail — define policies in natural language, group them into guards, and call them from any application or gateway.
TextGuard Lite
Fast text content guardrail covering violence, hate, privacy, jailbreaks, and 12+ harm categories. English and multilingual.
ImageGuard
Image safety classifier for harmful, unsafe, and risky visual content. Outperforms GPT-4o and Azure Moderation on standard benchmarks.
VideoGuard
Frame- and clip-level safety analysis for video content. Detects unsafe scenes, violence, sexual content, and deepfake-style manipulation.
AudioGuard
Audio safety classifier for speech and non-speech content. Catches abusive language, threats, sensitive topics, and voice-clone misuse.
CodeGuard
AI code vulnerability scanner for VS Code, Cursor, and Windsurf. Detects CWEs across Python, C/C++, and Java in AI-generated code.
What's common across the family
All VirtueGuard products share the same operational model so you can adopt multiple modalities without re-learning the stack:
| Aspect | Detail |
|---|---|
| Form factor | Small, low-latency classifiers that run alongside your application. |
| API style | REST endpoints (/api/textmoderation, /api/imagemoderation, /api/videomoderation, /api/audiomoderation, /api/topic_guard). PolicyGuard adds an OpenAI-compatible moderation endpoint. |
| Auth | API key (sk-vai-...) or JWT bearer token. |
| Deployment | SaaS, Docker Compose, Helm/Kubernetes, or Terraform/IaC for private/on-prem. |
| Output | Per-category probabilities, a top-line flag boolean, and optional reasoning. |
| Decision policy | Caller decides what to do when content is flagged: block, redact, alert, route to human review, or log. |
Choosing a guard
| If you need to… | Use |
|---|---|
| Enforce custom natural-language policies (brand, legal, domain-specific) | PolicyGuard |
| Filter text prompts, replies, or chat content | TextGuard Lite (or Multilingual) |
| Block unsafe images in uploads, generation, or moderation queues | ImageGuard |
| Screen video uploads, generated clips, or surveillance content | VideoGuard |
| Moderate calls, voice assistants, podcasts, or TTS output | AudioGuard |
| Catch vulnerabilities in AI-generated code | CodeGuard |
Where to deploy VirtueGuard
VirtueGuard is designed to sit on the request path of your AI system. Typical placements:
- Chat clients — input + output checks around every LLM call.
- Agent gateways — block policy-violating prompts, tool inputs, and model outputs before they leave the gateway.
- Upload pipelines — scan images, audio, and video on ingest before they reach storage or downstream models.
- RAG systems — guard user questions, retrieved snippets, and final answers.
- Pre/post-deployment review — pair with VirtueRed for continuous red-teaming and offline evaluation.
Next steps
- Pick the guard that matches your modality from the table above.
- Get an API key from the VirtueGuard dashboard (or
POST /api/api-keys). - Call the guard from your application — start with a single check on input, then add output checks once you've validated the behavior.
- For custom rules, layer PolicyGuard on top of the modality guards.