OpenAI Is About to Release Its First AI Model With ‘Critical’ Cyber Abilities
wired ·
OpenAI announced that its upcoming AI model Astra is the first to reach the company's threshold for "critical" cyber capabilities, meaning it can independently find and exploit previously unknown software vulnerabilities. The company paused some training workloads for Astra and a future model for several weeks to implement safety measures, and has now resumed work, planning to publicly release a version of Astra soon. However, at launch, the model's advanced cyber capabilities will be available only to select partners in OpenAI's Daybreak Blue early-access program, including Cisco, Cloudflare, and Palo Alto Networks, to help them shore up defenses. OpenAI is also adding a "misalignment monitor" to refuse unsafe queries, though it may occasionally flag legitimate activity and slow or pause user actions. The announcement follows similar incidents at Anthropic and Meta, highlighting industry-wide concerns about AI's hacking abilities. Astra reportedly outperforms other leading models on cybersecurity benchmarks such as ExploitBench and can chain multiple exploits to gain deeper access to target systems. Experts caution that while AI increases risks, existing digital security defenses and best practices remain durable.
wired