GPT-6 Cyber Built 948 Linux Exploits — Welcome to the N-day Exploit Era

Nine hundred and forty-eight working exploits for Linux kernel vulnerabilities. Eighty-one new Firefox flaws that Mozilla rated high severity. A remote code execution chain in Brotli that reaches Node.js downstream. OpenAI has raised GPT-6 Cyber's cybersecurity risk rating from High to Critical, and even after allowing for the usual self-reported benchmark inflation, the throughput is unlike anything we have seen from a single research team.

Every one of those findings is a candidate. The n-day exploit pipeline just got an industrial-grade front end, and the gap between patch and weaponisation is now measured in hours rather than weeks.

Bug harvesting at industrial scale

The headline numbers come from OpenAI's own harness: 1,724 Linux kernel vulnerabilities reproduced and 948 exploits developed, against 166 for GPT-5.6 Sol on the same setup — a factor of 5.7. On ExploitGym the model scored 48.5% versus 32.1%, and it completed 95% of tasks against known flaws in GitLab, PostgreSQL and Jenkins where GPT-6 Astra refused every one. Three hundred and four critical findings went to kernel maintainers with candidate patches attached.

We have learned to discount vendor benchmarks, and these deserve the same scrutiny — a model graded by the company that trained it is not an independent measurement. But the direction is unambiguous. Vulnerability research used to be a craft: one researcher, one bug, weeks of work. It is now a batch job, and the marginal cost of the next n-day exploit is approaching zero.

A deeper armoury for router and IoT botnets

Kernel bugs are the ones that matter to us. The devices that fill a botnet — routers, IP cameras, NVRs, set-top boxes, throwaway VPS — run kernels frozen years ago, and vendors patch them late or never. A working kernel exploit plus a userspace path is persistence on hardware nobody reboots. Mirai proved the model; every generation since has refined it, and a catalogue of hundreds of validated kernel exploits turns infection from a research project into a scanning loop.

Once the nodes are in place the traffic is the easy part: SYN flood, ACK flood, ACK-PSH flood, TCP handshake flood, UDP flood, UDP PPS flood, fragmentation flood, and UDP amplification through DNS, NTP, SSDP and CLDAP reflectors. The ip stresser and booter market resells that capacity by the terabit, ddos for hire panels quote PPS figures that would have been a world record a few years back, and the ddos service resellers downstream barely ask where the packets originate. Where we sit on that particular board is a matter of public record in our achievements.

Browser exploit chains and the L7 treadmill

Eighty-one new high-severity Firefox vulnerabilities, plus thirty existing issues that turned out to be worse than their original rating. Mozilla shipped fixes quickly, but the interesting part is the shape of the bugs: memory corruption in the engine, reachable from a page. That is the raw material for browser exploit chains, and it lands exactly when layer 7 defences have turned into a fingerprinting contest.

We already run browser emulation and headless browser fleets carrying plausible TLS fingerprints and JA3 fingerprints, rotating user agents, and clearing JS challenge and CAPTCHA gates where the implementation is lazy. The cloudflare bypass treadmill never stops. A fresh supply of client-side bugs is what turns those sessions from nuisance traffic into something that can compromise the visitor. On the flood side the classics still work — HTTP GET flood, HTTP POST flood, Slowloris, TLS flood, connection flood — and protocol-level flaws keep handing us new primitives. HTTP/2 Rapid Reset (CVE-2023-44487) went from disclosure to commodity in days. Expect the same compression here.

Triage is the new bottleneck

When discovery is cheap, the constraint moves downstream. Nine hundred exploits are worthless without a pipeline that clusters them by reachability, ranks them by target prevalence and automates the build-and-test loop across kernel versions and configurations. That is the work now: triage and exploit automation, not bug hunting. OpenAI's own red team showed the other half of the problem, chaining vulnerabilities to escape microVMs and land code on the underlying hosts, and Dragos found two zero-days in protective relay management software with almost no prompting. The skill floor for reconnaissance is dropping, and it is dropping for everyone.

Defenders get the same models, which is the only reason this remains a race rather than a rout. OpenAI's scope-adherence numbers — 99.93% of tests staying inside authorised boundaries — are precisely the claim we would want verified independently, because an agent that wanders off-task is an incident, not a benchmark.

So the economics of bug hunting have changed, and the n-day exploit window is the casualty. The teams that win the next twelve months will be the ones that automate triage as aggressively as discovery. If you want to see what that asymmetry looks like from the receiving end, our free stress test is open — point it at your own edge and watch the graphs at /contact.