OpenAI has launched GPT-5.6-Cyber, a new model built specifically for cybersecurity, and expanded Daybreak, its program that grants approved security defenders access to its cyber tools. The update cuts both ways: the model is at once more permissive and more capable for authorized offensive-security work — vulnerability research, exploit development, and penetration testing — yet it remains tightly gated. Only vetted users get the keys, which turns the product into a weapon the maker keeps under lock and key and hands out to licensed sheriffs.
The timing matters. AI-driven attacks are multiplying across industries, and the last few months have brought a string of security incidents involving autonomous AI agents. Reports describe agents that orchestrate intrusions, automate the discovery of weaknesses, and even deceive other systems into granting improper access. Each incident sharpens the same lesson: the machines doing the breaking are getting smarter at a speed defenders struggle to match. Against that backdrop, OpenAI is positioning itself as the defender's ally — while keeping the most aggressive pieces of its arsenal firmly under its own control.
GPT-5.6-Cyber is the direct successor to GPT-5.5-Cyber, released earlier as part of the company's push into specialized security models. What sets the new version apart, according to OpenAI, is the balance it strikes between freedom and accountability: the model can tackle more aggressive offensive-security tasks, from deep scanning to assembling functional exploits, but each capability unlocks only for users whose profile and approval level justify it. It is an architecture of granular permission, where the tool refuses to fire unless a legitimate context sits behind the request.
That approach puts OpenAI in the same lane as Google and Anthropic, which have both been building their own cyber-defense models. Each company charts a different route: some rely on usage policies, others on technical filters. OpenAI is betting on a third way — human and institutional vetting, with Daybreak acting as a professional credential for anyone who does offensive security for a living. Researchers, incident-response teams, and contracted testers make the list; the curious and the malicious do not.
Critics warn, however, that this revolving door is precisely where the risk hides. If a model this good at generating exploits falls into the wrong hands — through a leak, an approved user who gets compromised, or code that gets reused for malicious ends — the potential damage is far greater than before. And the sheer existence of such powerful offensive tools pressures the entire industry to speed up the attack-and-patch cycle, making zero-days an even more valuable commodity on the cybercrime market.
But the question hanging over the sector is no longer technical; it is one of trust and authority. Who, ultimately, decides who deserves to be a sheriff? Centralizing weapons this powerful concentrates enormous power in a handful of companies and turns trust into a scarce resource. Meanwhile, attackers ask no one for permission. OpenAI is putting the right tool in the right hands, yet the history of security has always been the history of who can better guard what should never have leaked onto the street. The next frontier is not building stronger defenses — it is deciding whom to share with the tools that could tear them down.
Sources: Axios, OpenAI, Business Standard, Digit
✓ Independent sources cross-checked and verified before publishing
