An AI That Can Break Into Software, and Why That Might Help Defenders

An original diagram of the loop security teams run: find a flaw, prove it is real, fix it before someone else uses it.

Imagine a locksmith who can test every door in a city in an afternoon. Wonderful if she's working for the city. Less wonderful if she isn't. That, roughly, is the question hanging over OpenAI's announcement on 5 November of a model called GPT-6 Cyber.

It's a restricted model built for security work: finding weaknesses in software, building working exploits to prove they're real, and suggesting patches. OpenAI rates it "Critical" for cybersecurity under its own Preparedness Framework, which is the company's way of saying this one needs real guard rails. Everything below comes from OpenAI's own figures as reported by The Decoder and CSO Online. I couldn't open OpenAI's pages directly, and nobody has published independent testing yet, so treat the numbers as claims.

What OpenAI says it did

Two results stand out. First, Firefox. OpenAI says the model found 81 previously unknown vulnerabilities, and that Mozilla accepted the reports and is shipping fixes. The two outlets describe the severity a little differently (one says critical, one says high-severity), so I'll just say serious.

Second, the Linux kernel, the core software under a huge share of the world's servers and phones. According to The Decoder, an automated system scanned roughly 5,000 security-relevant spots, shortlisted 2,308 candidates, and confirmed 1,724 real vulnerabilities inside isolated test machines. Working exploits were built for 948 of them. An earlier model had managed 166.

The bill for that run was about 1.2 million dollars in compute. That sounds huge until you remember what a single skilled human researcher costs for a year, and how many bugs they'd find.

A quick translation

An "exploit" is a recipe that turns a bug into a break-in. Finding a bug is like noticing a window that doesn't quite latch. Writing the exploit is climbing through it to prove the point. Defenders do this on purpose so they can fix the window before a burglar notices.

An original chart using OpenAI's own reported scores. ExploitGym is a test of remote exploitation skill. The honeypot test checks whether a model stays inside the systems it was allowed to touch.

The chart shows the jump on one benchmark, from 32.1 percent for the previous Cyber model to 48.5 percent. The smaller number matters just as much to me: in a honeypot test, the new model strayed out of scope in 0.07 percent of runs, against 5.51 percent before. A model that's good at breaking in and also good at staying in its lane is the combination you'd want.

Who gets to use it, and why I'm uneasy

Access is limited. OpenAI says GPT-6 Cyber starts with a select group of approved customers in its Daybreak Red tier, using the API and Codex, and opens up gradually over the coming weeks. It lists safeguards too: permission checks, monitoring, IP restrictions, and a recommendation to run it in isolated setups with humans watching.

Here's my honest reaction. The defensive pitch is believable. Software has more bugs than any team can find, and a tireless assistant helps. But the worry from outside experts, as CSO Online reports, is that finding problems can outrun fixing them. Fixes still need people to write, test and install patches, and one security commentator pointed to an OpenAI dashboard suggesting under a quarter of reported issues had been remediated so far. I haven't checked that figure myself.

There's also the point The Decoder raises: another group's results suggest similar skills may already exist in open-weight models that nobody controls. If so, gating one model only buys so much time.

And a fair caution on the marketing. OpenAI says the model found at least ten exploitation techniques that were new even to outside kernel experts. That depends heavily on how "new" is defined, and it's the company grading its own homework.

What to watch

Watch for three things: independent tests of these claims, how fast the Linux and Firefox fixes actually reach ordinary users, and whether other labs release comparable tools with fewer limits. For the rest of us the practical advice stays boring and good. Install updates promptly.

If software bugs can now be found at this speed, should the race be about finding them, or about fixing them? I'd love to hear what you think.

Sources: The Decoder, CSO Online, ThursdAI.