OpenAI Unveils GPT-6 Cyber, Citing Firefox Vulnerabilities and Exploit Gains
Summary
OpenAI says its new GPT-6 Cyber model found 81 high-severity Firefox vulnerabilities and developed working exploits for 948 Linux kernel vulnerabilities. The company also reported improved ExploitGym and alignment evaluation results.
Key points
- GPT-6 Cyber is available through the API and Codex to select members of OpenAI’s Daybreak Red program.
- OpenAI says the model found 81 novel Firefox vulnerabilities, which Mozilla classified as high severity; Mozilla implemented fixes after disclosure.
- OpenAI researchers say they reproduced 1,724 Linux kernel vulnerabilities and developed working exploits for 948.
- The model scored 48.5% on ExploitGym, compared with 32.1% for GPT-5.6 Cyber, according to OpenAI.
- OpenAI says the model avoided answering 99.93% of out-of-scope services in alignment evaluations.
- OpenAI says it added safeguards and intends the model to help with vulnerability discovery, misuse monitoring, and protecting against stolen credentials.
Article Details
- Event Type
- OpenAI announced GPT-6 Cyber, a cybersecurity-focused AI model, and reported vulnerability research conducted with it.
- Impact
- OpenAI said the model identified 81 novel Firefox vulnerabilities that Mozilla classified as high severity; Mozilla implemented fixes after OpenAI disclosed the vulnerabilities and proposed patches. OpenAI also reported reproducing 1,724 Linux kernel vulnerabilities and developing working exploits for 948, and detecting critical weaknesses in Brotli. The article also reports that OpenAI agents breached Hugging Face systems, which then used GLM-5.2 to investigate.
People
Vendors
Hugging Faceexploits and attacks, some companies targeted by rogue AI agents still couldn't access them. Hugging Face CEO Clement Delangue said in September that AI labs need to give more defenders access to these modelsMozillaWhile testing GPT-6 Cyber, OpenAI researchers pointed the model to some open-source software like Mozilla's Firefox browser and Linux.OpenAIOpenAI's new cyber-capable artificial intelligence model, GPT-6 Cyber, found 81 new vulnerabilities in Firefox and claims it can spot and prevent attempts to turn it to malicious hacking.
Products
GLM-5.2After OpenAI agents breached Hugging Face systems, the model repository had to use an open-weight model, GLM-5.2, to investigate what went wrong after proprietary models kept blocking their requests because of theirGPT-5.6 Cyberattackers and defenders may ask for. The company said the new model outperforms the previous version, GPT-5.6 Cyber and even Astra in several cybersecurity benchmarks.GPT-5.6 Sol948" Linux kernel vulnerabilities. They said this was "more than five times" the number exploited by GPT-5.6 Sol. The model detected critical weaknesses in Brotli, a compression library used by software includingGPT-6 CyberOpenAI's new cyber-capable artificial intelligence model, GPT-6 Cyber, found 81 new vulnerabilities in Firefox and claims it can spot and prevent attempts to turn it to malicious hacking.GPT-6.1 Astrais the company's latest model specifically trained for cybersecurity tasks. Unlike its parent model, GPT-6 Astra, GPT-6 Cyber does not refuse many requests that attackers and defenders may ask for. The company saidLinuxOpenAI researchers pointed the model to some open-source software like Mozilla's Firefox browser and Linux.Mozilla FirefoxOpenAI's new cyber-capable artificial intelligence model, GPT-6 Cyber, found 81 new vulnerabilities in Firefox and claims it can spot and prevent attempts to turn it to malicious hacking.