GPT-6 Cyber Release: Allows Writing Attack Programs

Just now, OpenAI released GPT-6 Cyber, a secure large model as a result: The model found 948 local privilege escalation vulnerabilities in the Linux kernel, as well as 81 high-risk vulnerabilities in Firefox

Other key points:

This model allows for writing attack programs
In OpenAI's Preparedness cybersecurity rating, this model is rated as Critical, the highest tier
Currently open to the first batch of approved Daybreak Red customers

Differences with GPT-6 Astra

GPT-6 Astra is capable of performing routine tasks well, but is blocked when carrying out attacks exploiting vulnerabilities, whereas GPT-6 Cyber is based on Astra and has been retrained with security tasks, thus loosening various security restrictions and achieving these results

Test	Exam Focus	Astra	Cyber
RemoteExploitBench	Remote attacks on known vulnerabilities	0% (Denied)	94.7%
ExploitGym	Exploiting specified vulnerabilities to breach targets	42.4%	48.5%
ExploitGym Hardened	Same as above, with added protections on target	29.7%	36.9%
ExploitBench Internal Version	Writing exploits for newly disclosed vulnerabilities	39.0%	42.0%
SEC-Bench Pro	Finding vulnerabilities in JS engines	85.4%	87.0%
SRE-Bench	Reverse analyzing binary programs	88.0%	92.8%
Web Penetration	Finding website vulnerabilities without source code (Recall Rate)	32.0%	40.3%

948 Linux privilege escalation vulnerabilities

In mid-October, OpenAI scanned the latest Linux kernel using an automated pipeline:

Filtering: Out of 40 million lines of code, approximately 2.8 million lines that regular users might encounter were selected
Analysis: 2,308 suspected vulnerabilities identified in 12 hours
Reproduction: Running test programs in an isolated virtual machine confirmed 1,724
Exploitation: Writing exploits, gaining root access on a standard desktop system with all default protections (SMEP, SMAP, KASLR) enabled, resulted in 948 after deduplication
Review: Researchers reviewed and reported the most severe 304 along with patches to the maintainers

On average, the API cost per exploit is $1,279. The hacker competitions Pwn2Own and TyphoonPWN historically awarded prizes of $20,000 and $70,000 for such achievements

Using the same process, GPT-5.6 Sol only found 166 in July, which means:

As models upgrade, more vulnerabilities will be discovered
For attackers with advanced models, everything is root

81 high-risk vulnerabilities in Firefox

Firefox has already been scanned by multiple generations of cutting-edge models this year, and GPT-6 Cyber found 81 previously unknown high-risk vulnerabilities, each with reproducible tests. Firefox 159, released on October 27, fixed 108 high-risk vulnerabilities, 41 of which were reported by OpenAI

Simulated power plant testing

The industrial control security company Dragos placed the Daybreak Red model in a simulated power plant, providing almost no hints. The model distinguished between the office network and power plant equipment, and found two unknown vulnerabilities (0days) in the engineering software that manages protective devices, thus gaining access to protective set values and circuit breaker tripping logic

This rollout simultaneously added interception: high-risk requests involving industrial control, basic services, and election systems will be blocked. Defense teams for power plants and water treatment facilities can apply for a separate Daybreak Red for Critical Infrastructure, with a funding plan of $1 billion for cash-strapped government agencies and basic service providers

How to use it

To use these models, one needs to apply for OpenAI's Daybreak program, which has two tiers:

Daybreak Blue: Uses the general model GPT-6 Astra, with some security restrictions eased, suitable for code audits, vulnerability checks, and writing fixes
Daybreak Red: Uses the specialized model GPT-6 Cyber, can write attack programs, perform authorized penetration testing, with stricter review processes. Companies need to implement single sign-on, multi-factor authentication, and IP whitelisting, while individuals must be invited, verified with real names, and use hardware security keys

However, if you just want to scan your own code, you can directly use OpenAI's tool Codex Security. After access to your repository, it will sequentially perform vulnerability discovery, issue validation, prioritization, task assignment, issue fixing, and checking, and then provide you with the final results

Discovery