Map / Model Security / WoolyAI
WoolyAI
woolyai.com ↗ · Open on the map
Model SecurityRuntime & Guardrails
Capabilities
Model capability monitoring for self-hosted open-weight LLMs. Detects in real time when a model starts generating restricted behavior (code, cyber/offensive activity, cryptography, tool calls) and enforces per-agent capability policy before the output reaches the agent and its tools.
Sensors / integration
Runs inside the customer's own inference runtime (e.g. vLLM serving Qwen behind an OpenAI-compatible endpoint, self-hosted or in a private VPC) and reads telemetry from the model's internals during generation, not just the prompt and response. Closed APIs (OpenAI, Anthropic, Gemini) are not supported because the runtime is not under customer control.
Protections
Per-agent policy on which capabilities a model may use, with Allow, Alert, Sandbox or Stop actions applied mid-generation, before generation completes and before the agent can act on the output.
Differentiator
Positions itself as 'model governance' rather than guardrails or agent governance: it decides whether the model may produce a behavior at all, using in-model signals instead of an external judge model or I/O filters. The tradeoff is that it only covers open-weight models the customer hosts. The founders previously built Qumranet/KVM, Ravello Systems (acquired by Oracle) and Veertu, and WoolyAI started as a hardware-agnostic GPU hypervisor (WoolyStack, CUDA abstraction for PyTorch across NVIDIA/AMD) before pivoting to AI model security in 2026.
Team & funding
- Founders
- Izik Eidus, Manisha Arora
Other Model Security companies
ActiveFence · Advai · Adversa AI · AI Aramid · Alice · AptaSentry · Arize AI · Attestable · Bosch AIShield · Citadel AI · Cranium AI · DataKrypto · See all →
Last updated · Get the “Listed on” badge