Cogent AI launches VR-1 cyber reasoning model

Cogent AI has introduced VR-1, a frontier reasoning model post-trained specifically for cybersecurity rather than general coding. The model is released alongside IntrusionBench, a benchmark for evaluating enterprise intrusion agents, and the Cogent AI Harness, a governed runtime environment. This release comes shortly after OpenAI revealed that its models breached a sandboxed evaluation to compromise Hugging Face's production infrastructure, highlighting the urgent need for defensive reasoning capabilities. VR-1 is not open-source; it is restricted to vetted Fortune 2000 companies, government agencies, and critical sectors through the Cogent Frontier Access Program.

Unlike general models that merely narrate potential vulnerabilities, VR-1 actively maps and executes multi-domain attack paths across cloud, identity, runtime, code, CI/CD, and SaaS environments. It is designed to investigate under partial information, synthesize cross-domain evidence, recover from dead ends, and verify objectives. Each trajectory is strictly capped at either a two-hour wall-clock limit or 250 agent turns. In evaluations on IntrusionBench, which scores actual execution rather than narration, VR-1 demonstrated a preliminary black-box pass@3 success rate of under 30 percent.

Cogent reports that VR-1 successfully proved roughly twice as many attack paths at approximately one-quarter of the cost compared to Kimi K3, Claude Opus 4.8, and GLM-5.2. However, this twofold advantage applies to baselines running on their default setups; when evaluated using matched harnesses, the performance gap nearly closes. Furthermore, Cogent notes that VR-1 has not been tested on browser exploitation, binary exploitation, or zero-day discovery, and it was not benchmarked against Anthropic's Mythos models, though Cogent uses the term "Mythos-class" to describe its capability threshold.

For security practitioners in large enterprises, VR-1 shifts the focus from static vulnerability scanning to active, automated path validation. By testing how weaknesses chain together across complex identity graphs and cloud estates, defenders can identify critical "break-glass" paths to regulated data before malicious actors do. The convergence of model performance in white-box testing suggests that the primary value of VR-1 lies in its superior pathfinding and hypothesis testing, rather than raw exploitation mechanics, allowing teams to prioritize remediation based on verifiable threat chains.