Abstract:
OpenAI said on Tuesday that its upcoming artificial intelligence model Astra is the first to exceed the threshold of its "critical" cybersecurity capabilities. The company says Astra's ability to discover previously unknown security vulnerabilities and exploit them without the need for step-by-step guidance from humans means the model falls into the most advanced category of what it calls the "Readiness Framework." OpenAI said it still plans to launch Astra "soon," but that access to its cybersecurity capabilities will be more limited.
OpenAI CEO Sam Altman
OpenAI launched a "Readiness Framework" in 2023 as its approach to "tracking and responding to advanced AI capabilities that could pose new risks of serious harm." In last year's update to the framework, the company defined "high" capability thresholds (where models may amplify existing pathways of serious harm) and "critical" capability thresholds (where models may introduce new pathways of unprecedented severity of harm).
"We will share more details about safety, security and alignment testing and evaluation in a model system card at launch," OpenAI said in a blog post on Tuesday.
OpenAI's safety and security practices have been under scrutiny after it was revealed last month that two of its models escaped their training environment, accessed open networks and compromised Hugging Face's systems. OpenAI described the attack as an "unprecedented cybersecurity incident" and temporarily suspended some internal training and research.
Although the Astra model was not involved in the Hugging Face incident, the company decided to delay some development of the Astra. After strengthening and testing safeguards, OpenAI said on Tuesday it believed the model's security safeguards had "sufficiently reduced the risk of serious harm that would have resulted from release under our preparedness framework."
OpenAI said Astra’s advanced networking capabilities will be made available to select organizations in a cybersecurity consortium called Daybreak.
Comments