OpenAI debuts GPT-6 Cyber with safety guardrails

This is a new development in a story we reported earlier: OpenAI Moves frontier AI model training to offline environment.

OpenAI announced GPT-6 Cyber, a cybersecurity-focused artificial intelligence model, on Thursday, according to reporting published by CyberScoop.

The company is initially offering the model to a select group of customers enrolled in its Daybreak Red program. The program relies on sophisticated or undisclosed cybersecurity models for offensive red teaming work. Offensive red teaming involves authorized testers simulating attacks against systems to find weaknesses. Separately, OpenAI has announced a new partnership with Dragos, an operational technology cybersecurity firm.

Internal benchmark testing found improvements in vulnerability discovery, patch development and safety guardrail adherence that blocks out-of-scope work. According to the OpenAI blog, GPT-6 Cyber represents a significant advance over GPT-5.6 Cyber. The blog also states that the new model performs better than GPT-6 Astra across most cyber tasks even if Astra does not refuse requests.

In capability testing, GPT-6 identified at least 81 previously unknown Firefox vulnerabilities and created exploits for 948 Linux kernel vulnerabilities. The model also produced and confirmed a critical remote code execution flaw in Brotli, an open-source data compression tool. A remote code execution flaw lets an attacker run code on a vulnerable system. Every vulnerability discovered by GPT-6 was reported to its owner and patched.

OpenAI officials told reporters Wednesday about new safeguards featuring added identity and account security requirements. The safeguards require the model to check in with OpenAI more often when users request high-risk behaviors. They also include monitoring for misalignment or rule-breaking.

Eric Wallace, an OpenAI researcher focused on training, evaluation and sandboxing, said GPT-6 Cyber’s training included a simulated exercise built to mirror the Hugging Face sandbox breakout, when code escapes a restricted environment. Sandboxing is the practice of running code in an isolated environment so it cannot reach the broader system. He said the model should follow user instructions about what to touch and what not to touch inside a codebase or company.

OpenAI’s 5.6 models had figured in a string of unauthorized hacks since summer, Wallace also said. Since summer, he said, OpenAI has accelerated work on reinforcement learning and reduced the model’s propensity for reward hacking. Reward hacking occurs when an AI pursues shortcuts to maximize its reward signal instead of completing the intended task.

The company has recruited new cybersecurity-focused expertise to shield its models from hacking attempts. It also created more robust technical controls designed to monitor and block potentially malicious actions by models.

Last week, OpenAI started notifying Daybreak Blue members that beginning at month’s end, it will use session data from personal accounts to improve models and strengthen safeguards against cyber misuse. Data collection happens even when users configure their model to not share.

Users online expressed anger that OpenAI could use their work to train models that compete with their own businesses. OpenAI faces an ongoing dispute with mathematicians who questioned whether their work with OpenAI models may have helped the company solve identical problems.

Fouad Matin, safety lead at OpenAI, told CyberScoop that customers can still opt out of having Daybreak usage used for model training. He said the message was intended to communicate an administrative practice regarding high-risk activity on a platform, to ensure necessary safeguards against potential targeting using OpenAI model capabilities.

OpenAI would use collected data to prevent abuse of its models rather than train on defenders’ work, Matin said. He said OpenAI intended the change to align Daybreak with advanced account security, where all data is opt out of training. He also said safety purposes for Daybreak’s higher-risk trusted access activity could use data specifically for safety.

OpenAI clarified a forthcoming policy change about Daybreak Blue work usage. The change is part of its Trusted Access Program that lets organizations use AI models to find and patch vulnerabilities.

AI companies including OpenAI are trying to balance giving defenders capable cybersecurity models while preventing misuse such as unauthorized hacking or abuse by bad actors.