OpenAI Halts Development of Astra Model Due to Cybersecurity Risks

OpenAI has temporarily suspended the development of certain components of its Astra artificial intelligence model. The decision followed an internal review revealing the system’s ability to independently identify vulnerabilities and launch attacks on real-world secured IT infrastructures. Under the company’s "Preparedness Framework," such capabilities require additional control measures.

In a blog post, OpenAI noted that preliminary tests confirmed Astra’s high performance level, which does not rule out the risk of reaching critical capabilities. However, the company emphasized that the model has not been used for unauthorized access, including in the recent Hugging Face platform incident*, where a data leak was caused by another unreleased model.

The Astra situation reflects growing concerns in the AI industry. In recent months, AI models have increasingly demonstrated potentially dangerous capabilities beyond test environments, sparking discussions among cybersecurity experts and regulators about the need for stricter oversight of AI developments.

OpenAI stated that it has implemented stricter security measures and paused internal work on Astra that does not meet new requirements. The company is also collaborating with government agencies and specialized AI safety organizations for comprehensive model testing.