Tibo Sottiaux, head of OpenAI Codex, stated that personal Daybreak Blue users must equip a qualified hardware key for network offensive capabilities starting October 1, while enterprise users are unaffected.

He specified that the hardware key is used to confirm the operator's ownership of the account and serves the strategy of "asymmetric acceleration of the defender." The personal path must enable advanced account security, allowing only FIDO2-compatible hardware keys for login, while software passwords and synchronized passkeys are not acceptable. Passwords, email, SMS verification codes, and email recovery will be disabled, sessions will be shorter, and conversations during this setting will not be used for model training.

Daybreak Blue is aimed at verified defense personnel, relaxing some system-level intercepts for general frontier models like GPT-5.6 Sol, covering vulnerability discovery, secure code review, malware analysis, incident response, and patch validation. Requests for dual-use, such as penetration of production systems, will still be rejected by this layer model. Higher privilege Daybreak Red requires separate approval, using GPT-5.6-Cyber, aimed at authorized vulnerability research, exploitation validation, and security testing.

Personal applications must be 18 years or older and use a qualified paid ChatGPT plan, completing identity verification; failures are typically non-appealable unless the company requests a resubmission. Access scope is bound to approved identities, workspaces or interface projects, models, and product surfaces. Blue does not automatically grant access to Red. The company previously noted personal hardware key nodes on September 1, now setting the execution date for network offensive capabilities to October 1.

Codex Security Cloud has default access to Blue layer network capability models, capable of scanning entire GitHub repositories, continuously reviewing new submissions, investigating and deduplicating findings, and preparing fixes for manual review, still operational when the notebook is closed, and appearing as a plugin on desktop and web. The company is also pushing Codex users from full release to automatic review, with high privilege actions assessed before execution, capable of intercepting destructive requests.

Buyers are verified enterprise security teams and critical infrastructure operators, while sellers are individual researchers who fail hardware key and real-name verification. The event-driven clarity: after the launch of offensive capabilities, personal access will be narrowed using physical keys and monitoring, leaving computing power and lenient protection for accountable entities. Beneficiaries include Yubico-like FIDO2 key vendors, enterprise security platforms, and OpenAI's own compliance narrative; those under pressure include independent white hats and personal accounts unable to maintain backup keys, losing keys without recovery codes may directly lose accounts.

The company also promises about $1 billion in subsidies within six months, aimed at state and local governments, critical infrastructure, community banks, non-profits, and open-source maintainers. Internal assessments indicate that GPT-5.6-Cyber has a completion rate of about 95% for authentication bypass, privilege escalation, and exploitation chain requests, compared to about 1.5% for GPT-5.6 Sol, about 2% for Blue path, and about 57.3% for the previous generation GPT-5.5-Cyber. The preparedness framework marks the network capabilities of both models as high, not yet reaching critical thresholds.

Source: Public Information