Models

OpenAI's GPT-6 Cyber found 81 unknown Firefox bugs and half a kernel's worth of exploits

OpenAI opened access to GPT-6 Cyber on November 5, a model built to hunt vulnerabilities and write exploits that prove they're real. The results from its first two campaigns are the part worth reading closely.

Daniel HarrisDaniel Harris
Heat: 1,580
Abstract steel-blue shield outline over scan lines

OpenAI opened access to GPT-6 Cyber on November 5, a model built to hunt vulnerabilities and write exploits that prove they're real. The results from its first two campaigns are the part worth reading closely.

What OpenAI GPT-6 Cyber found in Firefox

The Firefox campaign is the cleanest place to start, because those results came back with a second opinion attached. Mozilla's security team reviewed the reports and accepted them, then shipped fixes. That matters more than any benchmark score published alongside the launch.

Eighty-one of those findings were previously unknown severe vulnerabilities. That's the number OpenAI leads with. Thirty more were bugs Mozilla already knew about, which the model re-rated as more serious than the original triage suggested.

A model that finds a bug is interesting. A model whose bug reports survive a maintainer's review is a different category of tool, and that's the claim OpenAI is really making. Mozilla hasn't published its own account of how the reports compared with human submissions.

Inside the GPT-6 Cyber Linux kernel campaign

The kernel run is bigger and stranger, and it took an agent pipeline rather than a single prompt. OpenAI built a scout and analyst pair: one sweeps the source tree for security-relevant spots, the other digs into candidates and confirms them inside isolated virtual machines.

The sweep covered roughly 5,034 locations, about 7% of the codebase. That produced 2,308 candidates, which the pipeline narrowed to 1,724 reproduced vulnerabilities. For 948 of them, the system produced a working exploit that OpenAI says ran inside the isolated machine.

OpenAI compares that with GPT-5.6 Sol on the same tasks, which managed 166 exploits. Linux maintainers also received 304 critical findings with suggested patches attached.

Then there's the bill. OpenAI says the kernel run cost $1,213,430 in compute, about $1,279 per exploit. Pwn2Own and TyphoonPWN have paid $20,000 to $70,000 for comparable findings, so the price of finding a serious bug has moved in a way security teams should notice.

Kernel campaign

Result

Locations scanned

5,034, about 7% of the codebase

Candidates filtered

2,308

Vulnerabilities reproduced

1,724

Working exploits written

948

GPT-5.6 Sol exploits, same conditions

166

Why the OpenAI cybersecurity model stays inside guardrails

Guardrails are the part of this launch that decides whether the tool is usable by a company with a legal department. OpenAI trained the model to respect operational boundaries, and some of that training came from a honeypot inspired by the Hugging Face incident.

Without extra safety instructions, the model stayed in bounds in 99.93% of cases. Its predecessor managed 94.49%. A big jump.

Add safety instructions and OpenAI reports 100%.

Treat all three figures as vendor claims until someone outside the company checks them. No outside lab has published a replication, and boundary compliance is hard to test from the outside because the boundaries are defined internally, which means a model that fails quietly looks identical to one that never tried at all.

OpenAI also rates GPT-6 Cyber "Critical" in the cybersecurity category of its Preparedness Framework, the same rating GPT-6 Astra carries. A Critical label is the company's own ceiling warning, not an outside audit.

GPT-6 Cyber benchmarks come with an OpenAI label

Benchmarks are where the claims get specific, and every number in the table comes from the company that built the model. There's no independent test rig, no third-party replication, and no published test set you could rerun at home.

On ExploitGym, which measures how often a model turns a known weakness into a working attack, GPT-6 Cyber scores 48.5% against 32.1% for GPT-5.6 Cyber. GPT-6 Astra, the general model behind Daybreak Blue, reaches 42.4%. Astra with safeguards switched on scores 0.0%.

A second test covers web apps with no public source code to read. GPT-6 Cyber gets 40.3% there, against 32.0% for plain GPT-6 Astra. The gap is real by OpenAI's own measurement. Whether it holds under someone else's test setup is unknown, and that's the honest state of the evidence.

Who can actually use GPT-6 Cyber

Access is narrow by design, and that changes what the launch means for most readers. It starts with the API and Codex for selected Daybreak Red customers, and OpenAI says the group expands gradually.

GPT-5.6 Cyber opens to all Daybreak Red users at the same time, so most approved researchers get the older model first. Alpha access to GPT-6 Astra under Daybreak Blue also widens to more organisations and independent researchers.

There's no public signup, no consumer subscription and no published per-seat price. If you're an Android developer with a side project and a few thousand users, you can't buy your way in this week.

OpenAI lists the countermeasures customers should run: permission checks against a user's actual rights, monitoring for newly disclosed vulnerabilities, Private Safety Processing, IP-based access limits, and isolated environments with a human in the loop. It names power plants and water treatment as focus areas, which tells you who the company worries about.

Two programmes sit alongside the launch. Daybreak for Frontline Defenders funds security work at civil society groups, and a new Daybreak Red for Critical Infrastructure group handles authorised testing for utilities.

What GPT-6 Cyber means for Android and web developers

The useful takeaway for a normal developer sits in the price per exploit, because $1,279 changes who can afford to look for bugs in your code. What does that number buy someone with bad intentions? A week of scanning, roughly, instead of a consultant's invoice.

Epoch AI recently showed that an open model, GLM, could exploit known but unpatched Firefox vulnerabilities. Put cheap open models on one side and a frontier model that maps unknown ones on the other, and the window between a patch release and an attack gets shorter.

So update the browser on your test devices, keep your dependency tree current, and read the advisories for the libraries you ship. That advice is old. The reason to follow it now is that finding a bug got cheaper for everyone, including people who don't wish you well.

One more thing worth doing: check whether your app still ships a WebView old enough to miss recent patches. Android developers skip that one constantly, and it's the cheapest fix on the list.

GPT-6 Cyber is invitation only, and most readers will never touch it. The defenders who matter this month are the ones patching on the maintainer's schedule instead of their own.

Invitation only, for now.

Share This Story

Sources

Related AI News

A terminal window with a rising reward curve and a 27B badge beside it
Models

RadixArk Fine-Tuned a 27B Model on 1,200 Tasks and It Beat the Original Everywhere

RadixArk took an existing open model, spent 18 days and about 1,200 hand-built tasks on it, and changed nothing about its shape. The gains are real, and the absolute scores are still low.

Heat: 1,330
A laptop outline with three stacked layer sheets passing through a loop arrow inside it
Models

Inkling 2 Mini Fits on a Laptop by Running the Same Layers Three Times

Thinking Machines Lab built a model that thinks longer by running the same block of layers three times. The trick buys reasoning without extra memory, and the bill comes due somewhere else.

Heat: 1,710
A long teal audio waveform with a single orange accent bar on a dark field
Models

A 0.5B Speech Model Just Matched a 30B One on Real Phone Calls

Reson8 shipped a half-billion-parameter transcription model that gets within a tenth of a point of a 30B rival on noisy phone calls. The interesting part is what that does to a per-minute bill.

Heat: 560
Three neon rings orbiting a small glowing core on a deep blue field
Models

Gemini 4 Neon Lands in the App, and the Model Picker Loses Two Names

Google's fourth Gemini model in a month slots into the middle of the lineup, priced to be used every day. The app now asks subscribers to choose by effort, and two familiar names are gone.

Heat: 1,500