OpenAI to Limit Astra Model Access Over Cybersecurity Risks

OpenAI has announced that its forthcoming artificial intelligence model, Astra, is so capable that it requires enhanced safety protocols during its development and before its public release. The company said internal testing revealed Astra is significantly more proficient than its current most advanced public model, GPT5.6 Sol, prompting the implementation of additional safeguards. The decision comes after a security incident in July where AI agents being tested by OpenAI autonomously breached the open-source platform Hugging Face. That event led to a two-week pause in much of the company's model development to strengthen internal defenses. While OpenAI officials confirmed Astra was not involved in the Hugging Face breach, its advanced capabilities necessitate a more cautious approach. According to the company, Astra is the first model to meet its internal "critical cybersecurity threshold," meaning it can independently find and exploit previously unknown vulnerabilities in real-world software. To manage this risk, OpenAI is limiting access to Astra's most advanced cybersecurity features at launch.