Cogent uses GPT-6.1 Sol and GPT-6 Astra to automate vulnerability remediation at enterprises

- Reading
- 5 min
Cogent Security combines OpenAI models with its own knowledge graph to take vulnerabilities from discovery to verified remediation, without relying on a human analyst at every step.
The bottleneck isn't finding the flaw, it's fixing it
Any security team that has worked with a vulnerability scanner knows the problem: the tool generates thousands of alerts, but fixing each one requires understanding the affected system, who owns it, what dependencies exist, and whether the change could break something in production. This is the scope Cogent Security set out to tackle, according to material published by OpenAI (openai.com/index/cogent-security): the company builds what it calls "Cyber Defense Factories," defense factories that use AI agents to investigate, plan, and execute vulnerability remediation at enterprise scale.
The core idea isn't to generate yet another risk report. It's to have the agent carry the entire workload: from initial investigation to verified remediation, including confirmation that the risk was actually eliminated. For whoever maintains infrastructure, this shifts the question from "how do I prioritize alerts" to "how much of the remediation cycle do I trust to run without a human in the loop."
How the agent gains context (and permission) to act
The component underpinning this autonomy is what Cogent calls the Context Engine: a knowledge graph specific to each customer, with assets, vulnerabilities, identities, owners, and dependencies. Specialized agents continuously reconcile this graph with the security tools already installed at the company, so that investigation and remediation agents always see the same version of the environment.
Geng Sng, co-founder and CTO of Cogent, describes this as a "memory harness" for the company:
We don't have to relearn what the world looks like every time a new task happens.
Geng Sng, co-founder and CTO of Cogent
Reusing existing context, instead of rebuilding it with every call, reduces latency, inference cost, and inconsistency between runs. In practice, it's the same principle behind RAG applied to security infrastructure: the model reasons better when it doesn't have to rediscover the environment from scratch.
Access control is kept separate from reasoning. Agent Policies define what each company allows the agent to change and what requires human approval. The TRACE engine checks every tool call against these policies before execution, deterministically, without relying on another model inference to decide whether something is allowed. An Agent Gateway performs a second authorization check, and higher-risk work runs in an isolated sandbox. This design gives customers three possible levels of trust: delegate only investigation, require approval for any change, or allow autonomous remediation within defined limits.
The real test: the PHP2Shell flaws
The material cites a concrete case: when the critical flaws known as PHP2Shell were disclosed in PHP, customers who already had Cyber Defense Factories running in production used these agents to respond to the incident. GPT-6.1 Sol and GPT-6 Astra handled the investigation and remediation-planning reasoning, using customer context to identify affected systems, assess exposure, and build a plan that accounted for application dependencies and operational constraints.
From there, the agents orchestrated the fix following the approval and execution flow defined by each customer, and verified whether the vulnerability had actually been resolved. It's this complete cycle, from the public disclosure of the flaw to confirmation of the fix, that Cogent is selling as its differentiator compared to traditional scanners.
Numbers, and a customer that talks about the process
Cogent reports two concrete results: a Fortune 1000 automotive customer cut the average remediation time for critical vulnerabilities by 75%, and a Fortune 500 IT services customer reached 5 times more vulnerability triage capacity. These are figures provided by the company itself, without detailed methodology in the source, which calls for caution before treating them as a benchmark comparable across different environments.
Trading platform MarketAxess detailed how it uses the tool to decide what to fix first:
Cogent sees our environment the way an attacker would. It traces the route from the internet to a vulnerable system and checks controls on the way. Our remediation work now focuses on the greatest risks and the paths an adversary could use to break in.
Vijay Bala, Global CISO at MarketAxess
This detail matters more than the time-reduction figure: prioritization stops being based on isolated severity (CVSS, for instance) and starts factoring in the actual attack path within that specific environment.
Why GPT-6.1 Sol, GPT-5.6, and GPT-6 Astra, and not another provider
Cogent says it continuously evaluates open- and closed-source alternatives, and reports better reasoning on long security tasks, less hallucination in technical code, and more reliable execution with OpenAI's models. In practice, this means the models can chain tool calls in sequence and in parallel while keeping context across an entire remediation flow, without losing track of the investigation partway through.
The company segments usage by cost and complexity: GPT-6.1 Sol for general execution and GPT-5.6 for routine vulnerability-management tasks, where advanced reasoning isn't essential. Geng Sng also justifies the choice on operational cost grounds, which matters for anyone running multiple agents in parallel:
OpenAI models are more cost-efficient relative to the other providers. That's a big plus for us, especially if we're running these huge agent swarms.
Geng Sng, co-founder and CTO of Cogent
There's also an internal engineering gain: according to Geng, using OpenAI's APIs eliminated much of the code Cogent needed to build at the request orchestration and routing layer, reducing engineering overhead. Most of the company's engineers also prefer Codex for development, rather than using the same models in an open-source environment, because Codex already comes with tools the team wouldn't need to build from scratch.
What's left open
Cogent also uses privileged access to OpenAI's offensive research models, via Trusted Access for Cyber and a tier called Daybreak Red, for internal red-team research and attack-path analysis, on the logic that good defense requires understanding offense. The company's roadmap points toward increasingly autonomous remediation, covering everything from simple configuration changes to multi-step workflows that span several of a company's systems, alongside automated vulnerability discovery and validation as well.
For those who build and maintain infrastructure, the point of attention isn't whether the agent can write the right patch. It's the policy and sandbox design that decides what it can execute without supervision: TRACE, Agent Gateway, and Agent Policies exist precisely because giving a production agent, even a good one, autonomy to execute requires a deterministic boundary that doesn't depend on another model call to be trustworthy.
Translated from the Brazilian Portuguese original · Read the original
NaiveAI promises 2,100 tokens per second: does speed make up for code quality?
A hands-on test by Latent Space pitted the Chinese model Naive-N0.5-Flash against DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash. The speed promise holds up, but the trade-off in cost and code consistency shows up as soon as you look at the details.





