OpenAI Built an AI That Finds Hacks Before Hackers Do

OpenAI just gave a small group of vetted security teams access to an AI model built to hack like a professional, and to defend like one too.

GPT-5.6-Cyber is a version of GPT-5.6 Sol trained specifically for security work, exploit chains, privilege escalation, memory corruption, the stuff real attackers use. In OpenAI’s own internal testing it completed 95 percent of advanced cybersecurity tasks. The regular version of GPT-5.6 Sol, with normal safeguards on, completed about 1.5 percent. That gap tells you how much OpenAI had to loosen the leash to make this model useful for the job.

It is not open to just anyone. OpenAI is running it through a program called Daybreak, split into two tiers, Blue for general defensive work and Red for the specialized model itself, and companies have to apply and get vetted before they get access at all. Partners already in the program include names like Accenture, IBM, CrowdStrike, and Cisco.

OpenAI already put it to work. The model found two previously unknown vulnerabilities in Chrome’s V8 engine, the kind of bugs that could be chained together to break out of the browser’s sandbox. Those got reported and fixed instead of sold.

This is the uneasy trade AI security always makes. A tool good enough to find real vulnerabilities is also a tool that could cause real damage in the wrong hands. OpenAI is betting that keeping it behind a vetting wall, and pointing it at defenders first, tips the balance the right way.

Judging by the Chrome bugs it already caught, it is off to a decent start. If more labs start building tools like this and handing them to the people whose job is actually to protect us, that gap between attackers and defenders finally starts closing instead of widening.

——

Follow: @Ali Demi
Book your free AI clarity call, NOW!
https://buff.ly/TpWy277

——

Sources:
https://venturebeat.com/technology/openai-launches-gpt-5-6-cyber-with-reduced-refusals-95-completion-on-advanced-cybersecurity-tasks
https://www.axios.com/2026/08/10/openai-gpt-astra-restrictions-safety-hacking-defenders
https://dataconomy.com/2026/08/11/openai-expands-daybreak-with-new-gpt-5-6-cyber-model/

Repost this. Thanks.