OpenAI’s model helped identify a potentially catastrophic cybersecurity vulnerability in a programming language that runs millions of websites and applications. Greg Brockman, president of OpenAI, and Matthew Prince, CEO of Cloudflare — which helped protect websites before the flaw was publicly exposed — join Ed Ludlow on “Bloomberg Tech” to explain how AI is empowering cyber defenders and attackers.
--------
Like this video? Subscribe to Bloomberg Tech on YouTube:
/ @bloombergtech
Watch the latest full episodes of "Bloomberg Tech" with Ed Ludlow here:
• 'Bloomberg Tech' With Ed Lud...
What Greg Brockman said
Written from the verified transcript and checked against it. Every figure links to the moment it was said.
Greg Brockman, President of OpenAI, discussed the discovery of a catastrophic PHP vulnerability using OpenAI's models. He explained that a security researcher used GPT-5.6 Sol to find an initial version in August, and OpenAI's internal model autonomously found a more complete version in September. Brockman attributed this to AI's ability to handle vast context, contrasting with his own 2010 experience finding a vulnerability manually. He acknowledged that AI will help attackers exploit flaws, noting attempts are starting but not at massive scale. Brockman emphasized the 'defender's window' where frontier models stay ahead of attackers, and argued for societal processes, citing FBI court orders for botnet malware, as scales of vulnerabilities grow. He also touched on Dots adoption, saying it's 'really hockey sticking' and highlighted a friend's experience where Dots flagged a phishing email.
Key takeaways
OpenAI's internal model autonomously found a complete version of the PHP vulnerability in September, a month after GPT-5.6 Sol found an initial version.
Brockman estimates the Khalif researcher contributed 75-80% of the work using AI, while the OpenAI internal version was fully autonomous.
Brockman expects attacker exploitation attempts to ramp up, though they are not at massive scale yet.
Brockman called for societal processes, referencing FBI court orders for botnet malware, to handle growing vulnerability scales.
Dots adoption is 'really hockey sticking', with proactive use cases like flagging phishing emails.
0:00What is it that OpenAI's models actually did or do that allowed you to find something that I think researchers have not found for, like, years and years?
2:15Greg, as a follow-up, a specific question I got for you is how much of the work in this case study was literally done by AI versus a researcher?
4:14Greg, to the best of your knowledge, are there any examples of hackers using OpenAI's models to try and exploit this specific PHP flaw?
7:30Like, who decides that that should be the case, and where the line gets drawn?
10:19Is there any updated data you can give me on how Dots has gone from an adoption perspective and a user perspective?
AI helped uncover a potentially catastrophic vulnerability hiding for years in PHP. It's a programming language used to build and run millions of websites and web applications. OpenAI's technology helped researchers find the flaw. Cloudflare helped protect websites before that flaw was publicly disclosed. But AI could also help hackers exploit it faster. The concern is that the flaws could potentially let a hacker steal data or seize control of websites. Joining us now, OpenAI president Greg Brockman and Cloudflare CEO Matthew Prince. I'm grateful to both of you for your time. Greg, I want to get specific here. So what happened was, Khalif used GPT-5.6 Sol to find a version of this bug back in August, and then OpenAI independently kind of found a more complete version in September. What is it that OpenAI's models actually did or do that allowed you to find something that I think researchers have not found for, like, years and years?
First of all, thank you for having me. So you're exactly right. So Khalif had a security researcher use our models, very guided by that researcher's expertise and intuition. And a month later, we used our latest internal model that fully autonomously found this more complete version and very independently found this side... the kind of full version of what was released. And the thing to realize about the AI models is that, fundamentally, they're empowering tools and that they're able to work across a lot more context than people are. You know, I myself actually have found a security vulnerability way back in 2010 in an open source piece of software. And I remember finding a small function that looked like it wasn't quite implemented correctly, but all of the usages of it had been protected. But then actually someone added some code later that didn't protect in quite the right way. And so it's this kind of sleuthing that you do as a person, but it's very hard to keep all that context in your head, to have expertise across many different disciplines, and to think about how the software might be used in practice. And so that capability is... which is what makes AI so useful, right? That it means that it can help us solve problems. It's also something that helps us find flaws in the software that people have already written.
So for the Khalif side, I'd say... I don't know how to put an exact percentage on it, but the researcher had shown their deep expertise, let's call it 75%, 80%. Whereas for the OpenAI internal version, that was really... we set up an environment, and we took our most capable model with a lot of compute behind it, and that autonomously was able to find the full exploit.
Matthew, I want to go to you, and probably the scary bit, which was reading the blogs and the reporting. It seems like hackers tried to exploit these flaws before they were made public, right? Which raises the question whether AI was what found them first. Do you see what I mean?
Absolutely. And I think what we have to be clear about here is that AI is the real hero of this story. These flaws have existed for quite some time. They're extremely serious. And PHP, the underlying software that contains the flaw, is present and running on something like 70% of the world's websites. So this was very significant, and it was a very significant flaw. It's hard to know whether hackers, nation states, otherwise, had access to this flaw that had been present for years. And I think that what OpenAI did with their tools and what the researchers did was extremely responsible. They found a latent bug that was there. They found something that may have been exploited for many years. They responsibly disclosed it to the industry and companies like Cloudflare, and we worked tirelessly behind the scenes to make sure that our customers were protected and that the Internet was protected before it got out. But it's clear: it wasn't that AI created the bug. AI found the bug that was there, and it's incredibly important that we've now found it, fixed it, and the software is significantly more secure because of the work that OpenAI did.
