Artificial Intelligence · Cybersecurity
OpenAI unveils GPT-6 Astra and cyber safeguards
OpenAI said GPT-6 Astra had reached its Critical cybersecurity threshold, scored highly on exploit benchmarks, and would launch with restricted cyber functions alongside a new defender-focused programme.
- OpenAI unveiled GPT-6 Astra on Thursday and said it had reached its Critical cybersecurity capability threshold under its Preparedness Framework.
- OpenAI said Astra scored 100% on ExploitBench, compared with 78.5% for GPT-5.6 Sol.
- OpenAI said the release version would be limited to secure code review and patching and would refuse prompts for proof-of-concept exploits.
- OpenAI said Astra was rolling out to a small set of organisations and would later be available through ChatGPT Plus, Pro, Business and Enterprise, plus OpenAI API, Microsoft Azure and Amazon Web Services (AWS) Bedrock.
- OpenAI announced Daybreak for Frontline Defenders, a $1 billion project, and a pilot with U.S. Multi-State Information Sharing and Analysis Center (MS-ISAC).
OpenAI said on Thursday that it had unveiled GPT-6 Astra and that the model had reached the Critical cybersecurity capability threshold in its Preparedness Framework. The company also described it as its most intelligent and aligned model.
OpenAI said Astra scored 100% on ExploitBench, while GPT-5.6 Sol scored 78.5%. It also said Astra had performed better than GPT-5.6 Sol in tests using vulnerabilities disclosed from June to August 2026, including two zero-days in unnamed software.
The version being released was limited to code review and patching, and it would not comply with prompts for proof-of-concept exploits. OpenAI said it planned to widen access through OpenAI Daybreak in the coming weeks, with support for vulnerability validation, malware analysis and detection engineering.
Astra was rolling out to a small group of organisations and would later reach ChatGPT Plus, Pro, Business and Enterprise users, as well as the OpenAI API, Microsoft Azure and Amazon Web Services (AWS) Bedrock. OpenAI also announced Daybreak for Frontline Defenders, a $1 billion project for subsidised access, training and technical help, and said it had started a pilot with MS-ISAC.
Named in this story
Companies
- OpenAI
- unveiled GPT-6 Astra and announced the Daybreak initiatives
Organisations
- U.S. Multi-State Information Sharing and Analysis Center (MS-ISAC)
- the public-sector information-sharing body involved in a pilot
Products and systems
- GPT-6 Astra
- the new model being rolled out
- Preparedness Framework
- the framework under which OpenAI said Astra had reached the Critical threshold
- ExploitBench
- the benchmark on which Astra scored 100%
- GPT-5.6 Sol
- the previous frontier cyber-capable model used for comparison
- OpenAI Daybreak
- the planned access and safeguard programme for later expansion
- Daybreak for Frontline Defenders
- the global project aimed at giving defenders subsidised access and support
- Microsoft Azure
- one of the services that would host access to Astra
- Amazon Web Services (AWS) Bedrock
- one of the services that would host access to Astra
How the source tells it
The article read like a product announcement, mixing OpenAI’s own superlatives with urgent defensive language.
- vendor boosterism stacked superlatives and benchmark claims repeated the company’s own framing without independent challenge
- urgency the piece stressed an imminent rollout, coming safeguard changes, and a narrowing window for defenders
- threat framing it contrasted defenders with bad actors and used exploit and zero-day risk to heighten the stakes