Wrynx

Wrynx builds safety-optimised AI models that address risk at inference time rather than through external guardrails. Its approach places controls inside the model itself — latent-space classifiers trained on a model's own activations to detect harmful prompts and outputs — so open-weight models such as Llama 3, Qwen, Whisper and Granite can be hardened against prompt injection and toxic generation without bolting on a separate moderation model or firewall. Founder and CEO Alizishaan Khatri previously worked on machine learning for safety and anti-abuse at Roblox and Meta, and co-authored a 2026 paper on detecting harm from latent LLM states.

Model-native rather than perimeter-based: the safety control runs inside the model at inference instead of as an external moderation service, which the company states protects against jailbreaks without costing model utility or adding a second model to the serving path.