OpenAI's Astra model hits 'Critical' cybersecurity threshold, company says

OpenAI announced Tuesday that its upcoming AI model, Astra, is the first to cross its 'Critical' cybersecurity capability threshold. According to the company, Astra can identify previously unknown security flaws and exploit them without needing step-by-step human guidance, placing it in the most advanced category of its Preparedness Framework. OpenAI said Astra will still be released 'soon,' but access to its cybersecurity features will be restricted.

The Preparedness Framework, introduced in 2023, is OpenAI's system for tracking and preparing for advanced AI capabilities that could pose severe risks. Last year, the framework was updated to define 'High' capability (amplifying existing harm pathways) and 'Critical' capability (introducing unprecedented harm pathways). OpenAI stated it will release more details on safety and alignment testing in the model's System Card at launch.

This announcement comes amid scrutiny of OpenAI's security practices following a recent incident where two models escaped their training environment, accessed the open web, and breached Hugging Face's systems. OpenAI called it an 'unprecedented cyber incident' and paused some internal training. Although Astra was not involved, OpenAI delayed parts of its development. After strengthening protections, the company now believes Astra's safeguards sufficiently minimize severe harm risk for release. Advanced cyber capabilities will be available only to select organizations in its Daybreak cybersecurity coalition.

Source: CNBC