AI models & research
AI security models for local decisions.
Patronus builds specialized models for prompt injection, sensitive data, documents and agent actions. They run wherever AI is used, on workstations, servers and nodes.
5,000+
downloads per month
> 30,000
total downloads
6
open collections
01 · Model Zoo
Six models. One shared runtime.
Each model specializes in a concrete security task. Patronus Ark connects their representations and runs only the checks an input actually needs.
02 · NTDB
NTDBs decide when more depth is needed.
Non-Transformer Decision Blocks combine the shared model representation with small neural heads, LightGBM, logistic regression and class centroids. This produces an efficient first decision for each chunk.
Each chunk has two paths: if the fast NTDB result is sufficient, Ark keeps it. Only when additional context is likely to help does Lion Warden run the deeper transformer layers.
03 · Detection Engine
Three layers. One decision.
The pipeline combines different methods instead of sending every input through the same large model.
01
Heuristic detection engine
Deterministic heuristics and protocol-aware detection rules provide near-instant detection of known patterns with extremely low latency and no model inference.
02
Machine-learning detection
Gradient-boosted decision trees (LightGBM) evaluate structured risk signals from prompts, responses, tools and agent activity with high efficiency on endpoints.
03
Transformer security models
Compact transformer models trained specifically for security classification, with BERT-style architectures optimized for local deployment and real-time inference.
Security models for runtime protection.
Unlike general-purpose foundation models, Patronus models are built specifically for AI security, governance and runtime protection. The detection engine combines multiple analysis layers for fast and reliable decisions.
Ensemble architecture
Specialized detection layers connect heuristics, gradient-boosted trees, NTDBs and transformer models.
Low latency
Efficient checks decide immediately. More expensive model inference only runs for inputs that need additional context.
Local inference
Analysis runs on workstations, servers and nodes. Content does not need to be sent to another provider for the security decision.
Multi-Task AI
No single model is optimal for every security problem.
Patronus routes security tasks to specialized detection layers. Shared representations reduce memory and compute while individual heads remain trained for their specific task.
Reduced memory footprint
Multiple security tasks share central model components.
Ensemble decisions
Different signals combine into a more robust decision.
Intelligent routing
The Promote Gate sends only suitable chunks through the deeper transformer layers.
04 · FAQ
Frequently asked questions about the models.
What is Wolf Defender?
Wolf Defender is an open AI security model by Patronus for prompt-injection detection. It is available on Hugging Face and optimized for local inference.
What is an NTDB?
A Non-Transformer Decision Block uses the representation that has already been computed, combines it with small classifiers and makes an efficient first decision. Lion Warden does not need to fully process every chunk.
When does Lion Warden run?
The Promote Gate decides for each chunk whether more context is likely to improve the result. Only those chunks pass through Lion Warden's deeper transformer layers.
Does Patronus send content to the cloud for inference?
No. The models are built for local inference on workstations, servers and nodes. Content does not have to be sent to an external model provider for the security decision.
How are the models deployed?
Patronus Ark provides the shared runtime for Rust and Python. Patronus Protect uses the same technology as part of the finished desktop and runtime products.
Does Patronus use an LLM to detect AI?
No. Detection works with signals from network traffic and with small, specialized classification models such as Wolf Defender. No prompt is sent to a language model, and Patronus does not generate text itself.
Does Patronus send my prompts to OpenAI or another provider?
No. The check runs on the device. Only what your rules allow goes to the AI tool you chose yourself. No content goes to Patronus.
Do I have to replace ChatGPT or Copilot with something else?
No. You keep using your tools as before. Patronus sits on the computer in front of them and makes the usage visible, Protect also checks content and actions.
Is Patronus a VPN or a proxy?
No. Patronus runs as software on the device in the network path. Traffic takes no detour through third-party servers, and there is no central point that can fail.
Can my employer see what I write to the AI?
Content stays on the device. Only metadata can reach a central view: which tool, which category, which decision. Whether and how that metadata is evaluated is up to your company.
Does Patronus work with local models such as Ollama?
Yes. Detection works on how a connection behaves, not on the provider's name. Local models, internal services and new tools are covered just like the big providers.
Do I need an internet connection?
No. Detection, inspection and decision run offline on the computer. That includes isolated networks.
Does Patronus slow down my computer?
The models are built for their task: under 10 ms per detection and under 100 MB of memory. There is no noticeable difference in chat.
Research
Research behind the models.
Architecture, benchmarks and decisions behind Patronus Ark and its local model runtime.
How NTDBs Make Decisions Without a TransformerSep 12, 2026 · 10 minRead →How We Run Real-Time AI Security Inference On-Device on a CPUAug 31, 2026 · 14 minRead →Our AI-Security Model Zoo Is Now Open SourceJul 21, 2026 · 7 minRead →Wolf Defender v2: prompt injection detection that runs on the laptopSep 7, 2026 · 8 minRead →
All models on Hugging Face ↗