OpenAI's GPT-6 Cyber Can Hack Systems — and Patch Them Too

OpenAI just shipped an AI model whose entire job is to break into computer systems — and it's good enough at it that the company's own safety framework now rates it "Critical," the highest risk tier it has. The model is called GPT-6 Cyber, and unlike a chatbot that occasionally writes buggy code, this one is built to find real, working exploits on its own.

What GPT-6 Cyber actually does

Released on November 5, GPT-6 Cyber is a specialised version of OpenAI's latest model line, trained specifically for offensive and defensive security work. Offensive security here means actively probing a system for weaknesses the way an attacker would — not just describing them in theory, but writing and running the actual exploit code. OpenAI says the model can scan software for vulnerabilities, write working exploit code, and then turn around and suggest a fix for the same flaw.

The numbers behind the "Critical" label are what make this notable. On a benchmark called ExploitGym, which tests whether a model can successfully exploit a given known flaw, GPT-6 Cyber reportedly hit a 48.5% success rate, compared to roughly 32% for its predecessor. In testing, researchers reportedly used it to surface dozens of previously unknown high-severity flaws in Firefox, and to flag hundreds of critical issues in Linux code, complete with suggested patches sent to the people who maintain that software.

OpenAI's pitch is that this is a net positive: a model this capable, pointed at your own codebase before an attacker gets to it, finds the holes first. The company says it trained GPT-6 Cyber to stay inside the boundaries of whatever engagement it's given — not wandering off to attack systems it wasn't authorized to touch — and to report what it finds honestly rather than overstating results.

"A model that can find and exploit a zero-day" — an undisclosed flaw nobody has patched yet because nobody except the attacker knows it exists — "is only as safe as the guardrails around who gets to run it," is the blunt way one security researcher covering the release put it.

Why this matters beyond Silicon Valley

Here's the uncomfortable part: a tool built to automate finding and exploiting security flaws doesn't stay confined to the lab that trained it. The same reasoning that lets GPT-6 Cyber patch a Linux flaw is the reasoning that, in less careful hands, could automate attacks at a scale no human red team could match. That's precisely why OpenAI's Preparedness Framework — its internal system for grading how dangerous a model's capabilities are before release — bumped this one all the way to Critical, its top category.

For India specifically, this isn't an abstract debate. Microsoft's own threat intelligence has named India among the top targets globally for AI-assisted cyberattacks, with incidents now unfolding in minutes rather than the hours or days defenders used to have to respond. A model that can find exploitable bugs faster than human researchers raises the stakes on both sides of that equation — Indian banks, telcos, and government systems could use a tool like this to find their own weak points before criminals do, but the same capability, if it leaks or gets copied by a less scrupulous actor, speeds up the attacker's side of the race too.

This also lands just as India's own rules for AI incident reporting are taking shape. CERT-In, the government's cyber emergency response arm, recently introduced requirements for faster incident disclosure and kill-switch provisions for AI systems — rules written with exactly this kind of dual-use capability in mind. A security research model this powerful, operating across borders, is a direct test of whether reporting timelines built around traditional breaches can keep pace with AI-discovered ones.

What security teams should actually watch for

• Whether GPT-6 Cyber's exploit code or close variants start showing up in bug bounty submissions or real-world attack logs within weeks, which is usually how fast this kind of capability trickles into the wild.
• How OpenAI handles access — whether it's gated to vetted security firms and researchers, or available more broadly through its API.
• Whether Indian regulators or sector bodies like CERT-In or the RBI's cybersecurity wing issue any specific guidance referencing AI-discovered vulnerabilities, given the existing reporting framework.
• Whether rival labs — Anthropic, Google DeepMind — respond with their own dedicated security models, which would suggest this becomes a standard offering rather than a one-off.

The honest reading of GPT-6 Cyber is that it doesn't change what's possible so much as it changes who can do it and how fast. Finding a serious, unpatched flaw in a widely used piece of software used to take a skilled researcher days or weeks. If a model can do a meaningful chunk of that work in hours, the defenders who move first — patching before disclosure, disclosing before exploitation — win more often. The defenders who don't have access to anything like this, including plenty of smaller Indian firms running on tight security budgets, don't get that head start. That gap, not the model's existence itself, is the thing worth watching.