Cogent and OpenAI Build Autonomous Cyber Defense Factories

Introduction
Cogent Security is betting that the bottleneck in enterprise security is no longer discovery. In its framing, most organizations already generate far more vulnerability findings than their teams can realistically act on. The hard part is the remediation work that follows, which spans multiple teams and depends on knowing how an attacker could exploit a specific flaw, which business systems sit behind it, and which changes can be made without disrupting production.
Cogent's answer is a product category it calls Cyber Defense Factories: deployments inside a customer environment in which AI agents combine OpenAI models with enterprise-specific context and controls to carry a vulnerability from investigation through verified remediation. The company positions the factories as a way for defenders to operate at machine speed, responding to the accelerating pace of vulnerability discovery and exploitation. The architecture, the guardrails, and the early customer results are worth examining closely, because together they sketch what autonomous remediation looks like when it is actually pointed at production systems rather than a sandbox demo.
Context first, autonomy second
The central design claim from Cogent is that an agent is only as useful as its understanding of the environment it is acting in. To supply that understanding, Cogent runs a Context Engine that maintains an enterprise knowledge graph covering assets, vulnerabilities, identities, ownership, and dependencies. Specialized background agents continuously reconcile information across the customer's security tooling and other systems, so that the investigation and remediation agents share one consistent picture of the estate rather than each rebuilding it from scratch.
Geng Sng, Co-founder and CTO of Cogent, describes the result as a "memory harness for the enterprise." In his words, the team does not "have to relearn what the world looks like every time a new task happens." Reusing that accumulated context, Cogent says, cuts latency, reduces inference costs, and produces more consistent results across tasks. Upfront planning and heavier reasoning are handled by GPT-6 Astra, while GPT-6.1 Sol and GPT-5.6 drive the agents that assess exposures, execute permitted changes, and confirm that the risk has actually been eliminated.
Autonomy is bounded by explicit policy. Cogent's Agent Policies encode what changes each enterprise permits and which require approval, and a policy-enforcement engine called TRACE checks tool calls against those policies before execution. Those checks are deterministic and require no separate model inference, which matters for auditability: the gate does not depend on a model's judgment in the moment. An Agent Gateway independently verifies authorization, and higher-risk work can be run in isolated sandboxes. Customers choose how far to go, delegating investigation only, requiring human approval for changes, or permitting bounded autonomous remediation.
PHP2Shell as a live test case
Cogent points to the disclosure of critical flaws in PHP, the programming language behind a large share of the web, referred to in the source material as PHP2Shell, as an illustration of the model in practice. Customers running already-deployed Cyber Defense Factories turned to them when the flaws became public.
According to Cogent, GPT-6.1 Sol and GPT-6 Astra supplied the deep reasoning for investigation and remediation planning. Working through the context layer, the agents identified affected systems, assessed exposure, and built plans that accounted for application dependencies and operational constraints. They then orchestrated fixes through the customer's own approval and execution workflows and verified that the resolution held. The important detail for security architects is the sequencing: identification, exposure assessment, dependency-aware planning, policy-gated execution, and post-change verification, with the customer's workflows retained rather than replaced.
Reported outcomes and the reasoning-versus-cost trade-off
Cogent reports that customers are using the factories to resolve vulnerabilities faster and to expand investigation capacity. Among the cited results, a Fortune 1000 automotive customer reduced mean time to remediate critical vulnerabilities by 75%, and a Fortune 500 IT services customer reached five times its previous vulnerability-triage throughput. Global electronic trading platform MarketAxess uses Cogent to support a risk-based approach to exposure management. Vijay Bala, CISO at MarketAxess, said that Cogent brings together vulnerability data, business context, security controls, and attack path analysis to help validate real exposures and prioritize remediation where it matters most, and highlighted how quickly the Cogent team responds to emerging threats and converts feedback into practical capabilities.
Model selection is treated as an economic decision as much as a technical one. Cogent says it continuously evaluates closed-source and open-source alternatives and reports stronger reasoning on long-running security tasks, higher fidelity, fewer hallucinations on technical codebases, and more reliable execution with OpenAI models, which can chain tool calls sequentially and in parallel while retaining context across a remediation workflow. It routes GPT-6.1 Sol to general execution and complex reasoning, and GPT-5.6 to routine vulnerability-management work where advanced reasoning is less critical.
The engineering side effects are notable. Sng said a lot of the work Cogent previously had to do at the harness level is no longer required, and that OpenAI APIs have streamlined request routing and agent orchestration, reducing internal engineering overhead and accelerating time to market for complex security workflows. Most of Cogent's engineers also use Codex, which Sng said makes them more productive than running the same models in an open-source coding environment, largely because it ships with tooling the team would otherwise have to build. Cogent also receives technical support for internal red-teaming and attack-path research through Trusted Access for Cyber and Daybreak, with the Daybreak Red tier providing specialized models for authorized offensive and vulnerability research. As Sng puts it, good defense requires good offense.
What security teams should do now
Autonomous remediation is arriving through vendors rather than as an internal build, so the practical work for most teams is preparation and governance.
- Invest in context before agents. Ownership data, dependency maps, and asset inventory are the inputs that make agentic remediation safe. If that data is stale, an agent's blast-radius assessment will be wrong regardless of the model behind it.
- Write the policy, not just the prompt. Define explicitly which change classes are pre-approved, which need human sign-off, and which are never permitted. Deterministic enforcement outside the model is the control that survives an audit.
- Keep verification in scope. A closed ticket is not a fixed vulnerability. Require evidence that the exposure no longer exists before an agent's work is considered complete.
- Treat agent identities as privileged. Independently validate authorization checks, scope credentials tightly, and use isolated environments for higher-risk actions.
- Measure what the business cares about. Mean time to remediate and triage throughput are the metrics cited in Cogent's customer results, and they are the right baselines to capture before any pilot begins.
Grivyonx Expert Analysis
The most interesting thing about Cogent's design is not the model list, it is the separation of concerns. Reasoning lives in the models, context lives in a maintained knowledge graph, and permission lives in a deterministic policy engine that does not call a model at all. That split is what makes autonomous remediation defensible to a risk committee: when an agent makes a change, the organization can point to the rule that allowed it and the evidence that verified it. The unresolved questions are the usual ones for agentic security tooling. How does the context graph stay accurate as infrastructure churns, how are policy exceptions governed at scale, and what happens when an agent's plan is plausible but wrong in a way only a human would catch? The reported 75% reduction in remediation time and 5x triage throughput are meaningful, but they describe speed. The metric that will ultimately decide adoption is the rate of changes that had to be rolled back.
Conclusion
Cogent's Cyber Defense Factories represent a shift in where automation sits in the security workflow. Rather than adding another scanner to a queue that is already too long, the model is to close the loop from disclosure to verified fix, with enterprise context supplying the judgment and deterministic policy supplying the limits. Cogent says its roadmap includes increasingly autonomous remediation, which means the governance questions above will get harder, not easier, and the organizations that prepare their context data and change policies now will be the ones able to adopt it safely. As enterprises move agentic remediation from pilot to production, services such as Grivyonx's AI automation, DevSecOps and cloud security work are increasingly where those context, policy and verification requirements get engineered into the delivery pipeline.
- #Cogent Security
- #OpenAI
- #AI agents
- #vulnerability remediation
- #cyber defense
- #security automation
- #DevSecOps



