OpenAI Releases GPT-6 Astra With Its First Advanced Cyber Safeguards

OpenAI has released GPT-6 Astra to a limited set of customers, starting with participants in its Daybreak program for cybersecurity defenders. OpenAI says Astra is its first model to trigger enhanced internal protections under its Preparedness Framework because of its cyber capabilities. The launch puts those controls around a model OpenAI says can find new flaws and develop exploits without step-by-step human guidance.

Daybreak is the first gate

Astra is initially available to a limited customer group. OpenAI plans broader enterprise and consumer access in the following days, but Daybreak is its first route to users.

Daybreak separates approved security work into two lanes. Blue provides GPT-5.6 Sol with safeguards tailored to authorized defensive work. Red provides purpose-trained cyber models for authorized vulnerability research, exploit validation and security testing; OpenAI introduced GPT-5.6-Cyber through that lane.

The split is meant to address a practical overlap: vulnerability discovery, incident response and secure-code review can require knowledge that also enables system compromise. OpenAI assigns access levels according to the authorized work users are conducting.

Astra raises the computer-use bar

OpenAI describes Astra as a stronger system for autonomous computer use, software engineering and cybersecurity. It says the model can fill out spreadsheets and create websites, as well as find previously unknown security flaws and develop exploits across well-protected systems.

OpenAI also says Astra achieved a higher score than GPT-5.6 Sol while using fewer output tokens on ExploitGym, a cybersecurity test. That is a company-reported result on one benchmark, rather than a measure of performance in customer environments.

95.0%GPT-5.6-Cyber

OpenAI’s internal completion evaluation covers requests involving exploit-chain development, authentication bypass and privilege escalation.

1.5%Safeguarded GPT-5.6 Sol

The figures measure whether a model responds to advanced requests, not whether an attack succeeds or a defensive outcome is achieved.

The controls now have to work in use

OpenAI says it increased Astra’s cybersecurity protocols and added monitoring meant to rapidly detect and contain potentially misaligned actions. Daybreak also limits access to approved people and organizations conducting authorized work.

Daybreak’s stated operating controls

• Access uses identity verification, account security, monitoring, approved-use restrictions and legal attestations.
• Individual Daybreak accounts must adopt hardware security keys beginning September 1, 2026.
• OpenAI encourages Codex users to use auto-review mode, which reviews elevated-permission actions before execution and can block destructive requests.

Those measures govern access and tool use, but monitoring capable agents remains an open challenge. Chief scientist Jakub Pachocki said current observation methods may fail as models advance and potentially evade human monitors; he said OpenAI would withhold scaling if it could not regain sufficient confidence in alignment monitoring.

Astra also went through the White House’s voluntary vetting process for cutting-edge AI systems, according to OpenAI executives. Sam Altman said it was reviewed, while Greg Brockman said the administration requested no substantial safeguard changes. The next test is operational: whether restricted access and monitoring remain effective as Astra’s availability expands.