OpenAI Says New Model Meets Its ‘Critical’ Cybersecurity Threshold

OpenAI has announced that its upcoming Astra model meets its "Critical" cybersecurity threshold, according to a blog post released on September 1, following extensive testing of the AI model's capabilities. OpenAI's Preparedness Framework now considers Astra to possess the capability to identify unknown security flaws and develop methods to exploit them across numerous well-protected systems without guidance from a person.

In recent weeks, OpenAI has delayed Astra's development and launch to bolster and test safeguards against cyber misuse and unauthorized model actions. The company highlighted the recent incident in which its models breached AI platform Hugging Face, describing it as "an unprecedented cyber incident, involving state-of-the-art cyber capabilities."

Despite the breach, OpenAI believes its production safeguards at the time would have prevented the incident. The company has since implemented stronger safeguards for Astra, including training it to reliably refuse harmful cyber requests and respect safety restrictions.

OpenAI plans to make the Astra model available soon, but will impose limitations on access to its most advanced cybersecurity features. The article also touches upon the broader issue of corporate vulnerability, with PYMNTS noting that interconnected enterprises face challenges in incident response due to their interconnected networks.

The article cites research indicating that hackers are increasingly targeting middle-market companies, which often rely on third-party cloud providers, SaaS platforms, and managed service providers.