Cogent AI Releases VR-1: Cyber Reasoning Model Composing and Verifying Enterprise Attack Paths

Cogent AI released VR-1, a reasoning model post-trained specifically for cybersecurity, rather than gaining cyber capabilities as a side effect of general coding strength. VR-1 is not open-sourced; it is available only to vetted organizations through the Cogent Frontier Access Program, with guardrails, policy controls, and audit logging. The target market is large enterprises (Fortune 2000 and up), government, and defense, particularly in financial services, healthcare, SaaS, retail, telecom, and critical infrastructure. VR-1 investigates environments, tests hypotheses, and executes attack chains across cloud, identity, runtime, code, CI/CD, SaaS, and organizational context. The companion benchmark, IntrusionBench, grades execution, not narration, placing agents in controlled environments with execution-based verifiers. Cogent evaluated VR-1 across black-box, grey-box, and white-box settings; in white-box, models converge, suggesting VR-1's advantage is path-finding, not exploitation. Cogent reports VR-1 proves roughly twice as many attack paths at about a quarter of the cost (black-box pass@3) against Kimi K3, Claude Opus 4.8, and GLM-5.2. Cogent uses 'Mythos-class' as a scoped capability threshold and does not claim equivalence with Anthropic's Mythos models; VR-1 was not benchmarked against Mythos. Access is gated, and the model-agnostic AI Harness is the broadly deployable piece.