rvachev.orgEN / RU / 🤖
← Back to essays
· Essay · 1 min

OpenAI Released GPT-6 Cyber

OpenAI has released GPT-6 Cyber - a model for cybersecurity. The model is rated "Critical" and solves tasks for fixing vulnerabilities and writing exploits.

🛡 OpenAI has released GPT-6 Cyber - a model specifically for cybersecurity. And I think it's a good time to discuss where everything is heading.

According to OpenAI's own scale, the model has been rated "Critical" in cybersecurity, which is the highest risk level in their Preparedness Framework. Access is only available to verified organizations through the Daybreak Red program. In tests, the model solved 100% of vulnerability fixing tasks, 96% of exploit writing tasks, and 97% of red-teaming tasks. On the FrontierCyber benchmark with real software and hardware, it solved 90 out of 288 tasks, including zero-day vulnerabilities in browsers, phones, and cloud databases. Fully protected systems have not yet been breached.

Cybersecurity and software development are changing very quickly. I think we are not far from the moment when the idea of manually patching vulnerabilities will seem absurd. Attacking agents will work 24/7, and defending against them will also require agents. At a minimum, such an agent should be able to disable a system in which it finds a vulnerability. I would go further: defensive agents will constantly rewrite software to keep it secure in real-time. This means the code will become alive and will change constantly.

https://openai.com/index/introducing-gpt-6-cyber/

https://deploymentsafety.openai.com/gpt-6-cyber

#openai@rvnikita_blog #cybersecurity@rvnikita_blog #agents@rvnikita_blog #ai@rvnikita_blog #coding@rvnikita_blog

Newsletter

Get new posts by email

A weekly digest of what I write about AI, software and products. Unsubscribe anytime.