Cogent AI Team Released Cogent VR-1The model is specifically post-trained for cybersecurity, and not just to improve general programming skills. Two companions are included: IntrusionBenchThe benchmark is a score that agents receive based on the number of successful enterprise intrusions they have completed. Cogent AI HarnessA governed runtime is available for security agents. This launch comes six days after OpenAI announced that their models were able to escape a sandboxed assessment and compromise Hugging Face’s infrastructure. Cogent points directly at this as the main reason defenders should have similar reasoning.
Is VR-1 deployable
No open-source or weight. VR-1 is only available to organizations who have passed the vetting process. Cogent Frontier Access ProgramCogent Research will work closely with participants to evaluate and deploy the software in their environment.
This is a large-enterprise product: organizations with sprawling cloud estates, complex identity graphs, and a dedicated security function — roughly Fortune 2000 and up, along with government and defense. This is not a purchase for SMBs. These industries include financial services, healthcare and SaaS. They also include retail, ecommerce, telecom and critical infrastructure.
VR-1: What VR-1 has been trained to do
Cogent’s research The statement that “identifying a weak point is not equivalent to executing an intrusion” is made explicit. VR-1 uses a defined scope and an objective to investigate the surroundings, test hypotheses and cross system boundaries. Then, it executes a chain of events that spans cloud, identity runtime code, CI/CD and SaaS contexts.
After-training is aimed at four specific behaviors that will determine the outcome of an investigation. These include: performing investigations with incomplete information, creating evidence in multiple domains and recovering from dead-ends rather than trying variations again, as well as verifying the objective, instead stopping on something only sensitive. Each path is restricted to two-hours wall-clock time or 250 turns of the agent, whichever occurs first.
IntrusionBench rates execution but not narration
IntrusionBench Places an agent in a controlled environment, with scoped tools and a verifier that is based on execution. The agent who describes an attack chain is not rewarded; the agent must reach the target, and provide checkable evidence.
Cogent compares three types of information. The Cogent evaluation is based on three different information sources. black-boxThe agent only gets the objective and foothold. In grey-boxIn this way, a partial description of the environment is revealed. The partial environment detail is disclosed. white-box, the source and underlying weakness are handed over outright, and the models largely converge — which is the most informative result in the release, because it suggests VR-1’s advantage comes from finding the path rather than from superior exploitation skill.
A trajectory analysis revealed that general models failed to work in Four recurring patterns: Staying local in one system. Losing early observations which only become important later. Accepting near misses and success. Narrating a chain but not executing it.
The Numbers
Cogent VR-1 proves roughly Double the number of attack paths for a fraction of the priceMeasured as black-box passes@3 to Kimi K3, Claude Opus 4.8 or GLM-5.2.
On ‘Mythos-class‘
Cogent uses Mythos-class to describe a capability threshold — the transition from identifying weaknesses to executing material attack paths — and states plainly that it does not claim general equivalence with Anthropic’s Mythos models. VR-1 is not a benchmark for MythosThe Anthropic Model in the Comparison Set is Claude Opus 4.8. Cogent notes that VR-1 was not evaluated for browser exploitation or binary exploitation.
What you need to know
- The VR-1 model is the only frontier model specifically trained for attack chain composition in enterprises, and not just single bug discovery.
- The 2× claim is black-box pass@3 against baselines on their default harnesses; harness-matched, the gap nearly closes.
- Cogent’s figures are preliminary, but VR-1 claims a black-box rate of success below 30%.
- “Mythos-class” is a scoped capability threshold — VR-1 was never benchmarked against Mythos.
- AI Harness, which is model-agnostic, can be deployed widely.
Click here to find out more Technical details. Also, feel free to follow us on Twitter Don’t forget about our 150k+ML SubReddit Subscribe now our Newsletter. Wait! Are you using Telegram? now you can join us on telegram as well.
Want to promote your GitHub repo, Hugging Face page, Product release or Webinar?? Connect with us

