OpenAI announced GPT-6 Cyber on Thursday, a cybersecurity-focused model that the company says improves on vulnerability discovery, patch development and staying inside the boundaries a user sets for a job. GPT-6 Cyber will first be available to a limited number of customers in OpenAI's Daybreak Red program, which gives vetted teams more advanced or nonpublic cybersecurity models for offensive red teaming.
OpenAI officials said the model has found at least 81 novel vulnerabilities in the Firefox web browser during testing. It also developed and validated a critical remote code execution vulnerability in Brotli, a widely used open-source data compression tool. OpenAI tied the launch to its earlier work producing exploits for 948 Linux kernel vulnerabilities, and said all of the vulnerabilities were reported to their owners and patched.
In its announcement, OpenAI described GPT-6 Cyber as a major step up from GPT-5.6 Cyber and said it outperforms GPT-6 Astra on most cyber tasks, even when Astra does not refuse the request. The company illustrated the training flow it used to keep the model in scope during evaluations on ExploitGym.
Key facts
- 01OpenAI announced GPT-6 Cyber on Nov. 5, initially for a limited number of Daybreak Red customers.
- 02OpenAI says the model found at least 81 novel vulnerabilities in the Firefox web browser.
- 03The model also developed and validated a critical remote code execution bug in the Brotli compression library.
- 04New safeguards include extra identity checks, more frequent check-ins on high-risk requests, and misalignment monitoring.
- 05OpenAI also announced a partnership with operational technology security firm Dragos.
The access rules are tighter than for OpenAI's general models. Officials briefing reporters on Wednesday described additional identity and account security requirements, more frequent check-ins from the model to OpenAI when a request involves high-risk behavior, and additional monitoring for misalignment or rule-breaking.
Eric Wallace, an OpenAI researcher who works on training, evaluation and sandboxing for new models, said GPT-6 Cyber's training included a simulated evaluation meant to mimic the Hugging Face incident, in which models escaped internal sandbox testing and hacked into part of a major online AI code repository. The goal is a model that respects the limits of an engagement when it is turned loose on a real codebase.
“So we've been working on trying to make sure the model is really good at both listening to your instructions on what you should touch and what you should not touch inside a particular codebase or company.”— Eric Wallace, OpenAI researcher focused on training, evaluation and sandboxing
Wallace said that since the summer, when OpenAI's 5.6 models were involved in a series of unauthorized hacks, the company has accelerated work on reinforcement learning aimed at lowering the model's propensity for reward hacking. OpenAI has also brought in new cybersecurity expertise to protect its models and to build technical controls that monitor and block malicious model actions.
Alongside the model, OpenAI announced a partnership with Dragos, a cybersecurity firm that specializes in operational technology.
The launch comes as OpenAI adjusts data rules for Daybreak Blue, the defensive tier of its Trusted Access Program that gives organizations AI models for finding and patching vulnerabilities. OpenAI began telling members last week that, starting at the end of the month, data from Daybreak Blue sessions on personal accounts would be used to improve its models, including to strengthen safeguards against cyber misuse. Some users objected, saying they feared their work could help train competing models.
Fouad Matin, a safety lead at OpenAI, said the change was meant to align Daybreak with OpenAI's advanced account security terms and to allow data use specifically for safety purposes, given that Daybreak covers higher-risk activity. OpenAI clarified that customers can continue to opt out of sharing their Daybreak usage for model training.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
