Daily CSR

Cogent Security has unveiled Cogent VR-1, an advanced reasoning model purpose-built for cybersecurity and designed to function within live enterprise environments. Alongside the launch, the company introduced IntrusionBench, a benchmark that evaluates AI models against realistic enterprise attack scenarios. On this benchmark, VR-1 demonstrated its ability to start from a single point of compromise, independently identify viable attack paths, and achieve roughly twice the performance of other leading AI models.

While many current frontier AI models are capable of identifying isolated software vulnerabilities, real-world cyberattacks are far more complex. Attackers typically exploit a sequence of seemingly insignificant weaknesses spread across different systems. A minor flaw in a public-facing application, excessive user privileges, or an unverified software artifact may appear harmless individually. However, when combined strategically, these weaknesses can provide a pathway to an organization's most valuable assets.

VR-1 was developed to emulate the investigative approach of experienced cybersecurity professionals. Beginning with an initial foothold inside an enterprise, the model analyzes the surrounding infrastructure, correlates vulnerabilities across multiple platforms, and determines whether a complete attack chain is realistically achievable. It can also evaluate whether proposed remediation measures genuinely eliminate the risk or merely shift it elsewhere. As a result, security investigations that previously required weeks of expert analysis can now be completed within hours.

The launch comes at a time when autonomous AI-driven cyberattacks are becoming an operational reality rather than a theoretical concern. Recent incidents have shown AI-powered attacks targeting major government institutions and technology companies worldwide. Organizations therefore require equally sophisticated defensive AI systems to identify potential attack paths before adversaries can exploit them and to strengthen their environments against increasingly capable AI-assisted threats.

According to Vineet Edupuganti, CEO of Cogent Security, many security vendors have promoted AI-enabled solutions over the past several years, but VR-1 introduces a measurable standard for advanced security reasoning. He explained that the model operates directly within a customer's environment to identify the attack paths that a highly capable adversary could realistically exploit and then helps validate effective mitigation strategies. Since attackers are already leveraging similar capabilities, the company's objective is to ensure defenders have access to equivalent technology within a controlled and well-governed framework.

Addressing the Limitations of Existing AI Models
Traditional AI security benchmarks primarily assess whether a model can exploit an individual vulnerability in isolation. IntrusionBench instead evaluates the model's ability to conduct realistic, multi-stage enterprise attacks. In the benchmark, the AI agent begins with limited access and a defined objective, then attempts to navigate through cloud infrastructure, identity management systems, and internal enterprise applications to reach a specified target. Performance is measured based on successfully executing verified attack chains rather than simply identifying theoretical possibilities.

Developed by Cogent Security, IntrusionBench demonstrated that VR-1 performs particularly well when operating with minimal prior knowledge of the target environment. As progressively more environmental information becomes available, competing models improve and the performance gap narrows. Once the complete source code and underlying vulnerabilities are disclosed, the models perform at similar levels, suggesting that VR-1's primary strength lies in investigating uncertainty and connecting fragmented information rather than merely exploiting known weaknesses.

Geng Sng, Co-founder and CTO of Cogent Security, noted that existing frontier AI models can identify attack paths largely because of their general coding and reasoning abilities, rather than being specifically engineered for cybersecurity. VR-1, by contrast, has been trained explicitly to understand the complex, multi-stage attack sequences commonly encountered in enterprise environments, enabling it to outperform general-purpose AI models on practical defensive security challenges.

Secure Runtime Environment
In addition to VR-1, Cogent introduced the Cogent AI Harness, a runtime environment designed to safely deploy AI reasoning models within operational enterprise networks. The platform provides controlled access to environmental context, policy-aware tools, governance mechanisms, and independent verification processes. Every action performed by the AI is validated against organizational security policies and supported with verifiable evidence.

Cogent follows a validation philosophy described as "prove, don't detonate," meaning that vulnerabilities are confirmed without causing operational disruption, establishing persistence, or damaging production systems.

Controlled Availability
Recognizing that advanced cybersecurity capabilities could also be misused, Cogent has chosen not to release VR-1 as an open model. Instead, access is restricted to vetted organizations through the Cogent Frontier Access Program, where customers receive the model with built-in safety controls, governance policies, and comprehensive audit logging.

Organizations accepted into the program collaborate directly with Cogent's research team to evaluate the model, validate its performance within their own environments, and deploy it securely. Eligible participants can also receive a Frontier Model Risk Assessment, which identifies the attack paths an advanced AI model could realistically exploit, highlights vulnerabilities that become more significant as AI capabilities evolve, and recommends opportunities to safely leverage AI to strengthen organizational security.

Organizations can apply here.