AI development continues to accelerate across software engineering and cybersecurity, with coding startup Cognition reportedly exploring a new funding round that could value the company at $40 billion, while OpenAI has launched a new purpose-trained model aimed at advanced cyber security tasks.

Cognition AI is reportedly in early talks with investors over a funding round that could increase its valuation by more than 50%, less than three months after the company closed a $1bn raise at a $26bn valuation.

The startup’s annualised revenue run rate is now nearing $1bn, roughly double the level it had reached at the time of its previous funding round.

Any new raise would be expected to bring in more than $1bn, although Cognition could ultimately decide not to proceed or may seek different terms, according to a source cited by Bloomberg.

Founded in 2023, Cognition has built its reputation around Devin, an AI agent designed to carry out engineering tasks rather than simply suggest code.

Devin can read through a codebase, develop a plan, write code, run tests and attempt to correct its own mistakes with limited developer oversight. Cognition describes the technology as an “AI software engineer”, rather than a coding assistant.

Cognition’s valuation has risen rapidly over the past year. In August 2025, the company closed a $500m round led by Peter Thiel’s Founders Fund that valued the business at close to $10bn.

That funding came weeks after Cognition acquired rival start-up Windsurf, following the departure of Windsurf’s founders and research leads to Google as part of a $2.4bn licensing deal.

Cognition’s current backers include General Catalyst, Khosla Ventures, Pear VC and 8VC, alongside Founders Fund.

A valuation of $40bn or more would also take Cognition beyond the $29.3bn private-market valuation previously secured by Cursor before SpaceX agreed to acquire the company. Cursor had been in talks to raise funding at a $50bn valuation in March after reaching $2bn in annualised revenue a month earlier.

OpenAI Expands Cybersecurity Models

Meanwhile, OpenAI is expanding its own specialised AI capabilities with the launch of GPT-5.6-Cyber, a large language model purpose-trained for cybersecurity tasks and based on its GPT-5.6 Sol frontier model.

The company has also restructured its Daybreak programme with the introduction of two access tiers, Daybreak Blue and Daybreak Red.

Daybreak Blue is designed for defensive cybersecurity work and will provide members with access to frontier general-purpose models including GPT-5.6 Sol, with what OpenAI described as “system-level safeguards” for authorised security work.

Supported use cases include vulnerability discovery, secure code review, malware analysis, incident response and patch validation.

Daybreak Red will provide access to specialised cyber models including GPT-5.5-Cyber and the new GPT-5.6-Cyber, supporting more advanced defensive and offensive security tasks such as vulnerability research, exploit validation and security testing.

In a blog announcing the changes on 10 August, OpenAI said GPT-5.6 Sol “delivers state-of-the-art performance on cybersecurity tasks.”

However, the company has also implemented safeguards intended to “screen cybersecurity-related requests to prevent misuse,” acknowledging that these measures could “also block legitimate defensive work.”

Those additional guardrails will remain in place for general users of GPT-5.6 Sol but are removed for Daybreak Blue members.

OpenAI said that even without the additional restrictions, GPT-5.6 Sol will continue refusing some “highly dual-use cybersecurity prompts (e.g. pentesting production systems).”

GPT-5.6-Cyber, which is only available through Daybreak Red, is intended to provide greater capability for advanced security work.

According to OpenAI, the cyber-focused model reduces refusals while improving its performance on sensitive cybersecurity tasks.

The company said GPT-5.6-Cyber completed 95% of a set of sensitive requests covering areas including exploit-chain development, authentication bypass and privilege escalation.

Recommended

• Seven In Ten UK Employees Using Unapproved AI Tools, Microsoft Finds
• Report: AI Adoption Elevates Human-centric Skills
• Will AI Enable Workers to Secretly Juggle Multiple Jobs?

By comparison, general-access GPT-5.6 Sol completed 1.5% of the same tasks, rising to 2% when used through Daybreak Blue. OpenAI’s previous specialised model, GPT-5.5-Cyber, completed 57.3%.

OpenAI also compared GPT-5.6-Cyber against general-access GPT-5.6 Sol, the Daybreak Blue version of GPT-5.6 Sol and GPT-5.5-Cyber across cybersecurity-specific benchmarks including ExploitGym and ExploitBench.

The evaluation also covered areas including zero-day vulnerability discovery and vulnerability reporting, with OpenAI saying GPT-5.6-Cyber outperformed its other models across all of the tested tasks.

The company said it has also used GPT-5.6-Cyber to identify CVE-2026-15903, a high-severity vulnerability affecting V8, Chrome’s JavaScript engine.