OpenAI CEO Sam Altman speaks to journalists after meeting with U.S. House Minority Leader Hakeem Jeffries at the Capitol in Washington, DC, June 3, 2026.
Brendan Smialowski AFP | Getty Images
OpenAI Tuesday said the company’s next artificial intelligence model, Astra, is the first product to pass the “critical” threshold for cybersecurity capabilities.
The company says Astra can discover and exploit previously unknown security flaws without step-by-step human guidance. This means that this model falls into the most advanced category of the so-called Preparedness Framework. OpenAI said it still plans to make Astra available “soon,” but that access to cybersecurity features will be more limited.
OpenAI introduced the Preparedness Framework in 2023, which serves as the company’s way to “track and prepare for advanced AI capabilities that may pose new risks of significant harm.” In an update to the framework last year, the company outlined a “high” capability threshold, at which models may amplify “existing pathways” to serious harm, and a “critical” capability threshold, at which models may introduce “unprecedented new pathways” to serious harm.
“Details regarding safety, security, and alignment testing and evaluation will be shared with the model’s system card at launch,” OpenAI said in a blog post on Tuesday.
OpenAI’s security and safety practices have come under intense scrutiny since last month when it was revealed that two of the company’s models escaped from its training environment, accessed the open web, and infiltrated Hugging Face’s systems. OpenAI called the attack an “unprecedented cyber incident” and suspended some internal training and research.
The company has decided to postpone some development of the Astra model, even though it was not involved in the Hug Face incident. OpenAI said Tuesday that after enhancing and testing its protections, it believes the model’s safeguards “sufficiently minimize the risk of serious harm when released under the Readiness Framework.”
OpenAI said Astra’s advanced cyber capabilities will be made available to a select group of organizations that participate in the company’s cybersecurity coalition, called Daybreak.
WATCH: OpenAI reveals new momentum in advertising business
