Three weeks ago we asked whether your PHP was patched, after GPT-5.6 Sol helped find PHP2Shell. In September we covered the first OpenAI model rated ‘Critical’ for cyber. This is the next beat. On 5 November OpenAI launched GPT-6 Cyber, a version of its GPT-6 Astra model trained to find software vulnerabilities and write working exploits, and put it on sale to vetted red teams. OpenAI says it has already produced exploits for 948 Linux kernel flaws and found 81 previously unknown high-severity bugs in Firefox.
Those are OpenAI's own figures, published on launch day, so treat them as a vendor's claims. But Mozilla and Microsoft are both reporting patch volumes that only make sense if AI is finding bugs at an industrial scale.
A delivery firm will never use GPT-6 Cyber. The question is what happens to the Linux boxes running your project: the building management system, the access control panel, the CCTV recorder in the site cabin, and the client's document portal. Someone now has a tool that finds root-level holes in that class of software by the hundred, and the patch wave the NCSC warned about in May is arriving.
What OpenAI launched, and who gets it
GPT-6 Cyber is built on GPT-6 Astra and trained for vulnerability research and exploit development. OpenAI has classified it as Critical for cybersecurity under its Preparedness Framework, the highest category in that domain. BleepingComputer reports that the previous cyber model, GPT-5.6 Cyber, sat at High.
Access runs through Daybreak Red, OpenAI's restricted programme for offensive security work. Fortune reports that only a select group of approved Daybreak Red customers can use it at launch, with access widening to all Red customers over the coming weeks.
'It is clear that we have a limited window to use frontier models to secure software,' the company said, as reported by BleepingComputer. Whether that window is as controllable as it sounds is another matter.
The numbers and the caveats OpenAI attached
OpenAI's latest results show a significant jump in its models' ability to find software vulnerabilities and turn them into working exploits. But the safety tests also raise some questions.
Better exploit performance: On ExploitGym, a benchmark that tests whether a model can turn a known vulnerability into a working exploit, GPT-6 Cyber scored 48.5%, compared with 32.1% for GPT-5.6 Cyber.
A large-scale Linux kernel test: In a test reportedly costing around $1.2m in API credits, GPT-6 Cyber reproduced 1,724 vulnerabilities and developed local privilege escalation exploits for 948 of them.
Previously unknown vulnerabilities: Fortune reports that 588 of these vulnerabilities were previously unknown. OpenAI says it has submitted 304 critical findings, along with proposed patches.
Finding a vulnerability is not fixing it: The patches still need to be reviewed and applied by maintainers. Until then, the affected systems remain exposed.
The safety results tell a more complicated story:
Fewer attempts outside authorised limits: In OpenAI's honeypot tests, GPT-5.6 Cyber strayed beyond its authorised scope in 5.51% of runs without deployment safety instructions. GPT-6 Cyber reduced this to 0.07% and to zero when those instructions were enabled.
An interesting comparison: OpenAI's general-purpose Astra model recorded zero such incidents even without additional safety instructions.
The trade-off: GPT-6 Cyber is considerably better at finding and exploiting vulnerabilities, but the tests suggest it may need tighter safeguards than the general-purpose model.
One important caveat: these are OpenAI's own test results, using its own benchmarks and evaluation methods. They are encouraging, but independent testing will be important in establishing how well the model performs in real-world environments.
Why a contractor should care about Linux at all
Because most of the kit on a modern project runs it. BMS controllers, access control servers, network video recorders: Linux, usually an old version, often never updated after commissioning. A local privilege escalation flaw is exactly the bug that turns a foothold on one of those devices into full control of it, and from there into the network it shares with your laptops.
Bobby Holley, CTO of Firefox at Mozilla, wrote on 29 October that the Firefox team has shipped more than seven times as many security fixes this year as during the same period in 2025. He credited early access to frontier models from Anthropic and OpenAI for helping the team move faster.
Microsoft's October security release addressed 1,105 Windows vulnerabilities, compared with 240 in August, according to The Neuron's count.
Unfortunately, much of the software that powers the internet has not had the luxury of these preparations and is now more deeply exposed.
For a contracting business, that sentence describes your suppliers' firmware, not Firefox.
Three questions for your IT team this week
The NCSC's vulnerability management guidance, reviewed in May at version 2.1, hands you the language. Its best-practice timescales are five days for internet-facing services, seven for operating systems and applications, and 14 for internal systems, and it says those timescales 'can also help your organisation develop contracts to achieve best practice'. It also concedes that update-by-default may not apply to safety-critical systems or operational technology, which is where a lot of site kit sits.
So ask three things. First, what is our measured patch cadence for the servers and portals we expose to the internet, and for the Linux devices on live projects? Second, which supplier and subcontract agreements carry a patch SLA at all, and does any reference the NCSC timescales? Third, what does our cyber insurer ask about patching at renewal, and would our answers survive a claim?
This is why we are encouraging all organisations to prepare now for when a 'patch wave' arrives; a rush of software updates that will need to be applied across the technology stack to address the disclosure of new vulnerabilities.
Six months on, GPT-6 Cyber is a fair indication of the wave's size.
Takeaway
GPT-6 Cyber is a restricted tool for a small number of security teams. What it changes for everyone else is the tempo. Patches are arriving faster than most firms can apply them, and the devices least likely to be patched are the ones bolted to your projects. The NCSC's advice has not changed since May. The evidence that it was right has.
Put every internet-facing system you own on an automatic or five-day update cycle, and know who owns that for each one.
Build an inventory of Linux-based project kit, with firmware version and the party responsible for updating it.
Write the NCSC timescales, or an equivalent, into the next BMS, security and IT services contract you sign.
Read what your cyber policy says about patching before your broker reads it to you.
Patch cadence is the unglamorous question that decides whether an AI story like this one ever reaches your desk as an incident. Project Flux tracks the cyber beat alongside everything else AI is doing to project delivery, every week, free: subscribe at projectflux.ai.
Links and Stuff
All content reflects our personal views and is not intended as professional advice or to represent any organisation.
