Cogent Security Introduces VR-1 Cybersecurity Reasoning Model

Cogent Security announced in a press release Cogent VR-1, a frontier reasoning model trained for cybersecurity tasks inside live enterprise environments. The San Francisco company also released IntrusionBench, a benchmark that tests whether an agent can move from an initial foothold to a defined target across cloud infrastructure, identity systems, and internal tools.

In Cogent's benchmark results, VR-1 proved twice as many enterprise attack paths as Kimi K3, Opus 4.8, and GLM 5.2, at about one quarter of the cost. The hardest IntrusionBench setting gave the agent no information about the environment at the start. The gap narrowed when more environment details were disclosed, and models converged when source code and the underlying weakness were shown.

VR-1 is designed to map an environment from a foothold, connect weaknesses across systems, verify reachable attack paths, and check whether a proposed fix closes the path. The system can run this work in hours rather than weeks for some tasks.

Cogent also introduced the Cogent AI Harness, a runtime for deploying security agents with environment context, scoped tools, policy enforcement, verification, audit logging, and customer controls. VR-1 will not be released openly. Access is limited to vetted organizations through the Cogent Frontier Access Program, where participants can evaluate the model and receive a Frontier Model Risk Assessment.