OpenAI halts Astra projects due to risks of cyber attacks

Development of model will now take place in strictly isolated sandbox environments to prevent autonomous attacks

Pro

Sam Altman, OpenAI. Image: Shutterstock

OpenAI has implemented stricter safety measures and halted certain internal projects after it was unable to rule out that its next-generation model, Astra, possesses “critical” cyber security capabilities.

According to the organisation’s established safety framework, a model is classified as critical if, without human guidance, it can independently carry out complex attacks on secured infrastructure or discover and exploit previously unknown software bugs, also known as zero-day vulnerabilities.

Recent internal assessments and evaluations by external specialists indicate that Astra may be capable of independently carrying out highly advanced cyber operations. OpenAI has therefore tightened its security protocols and suspended all internal work related to Astra that does not comply with these updated requirements.

To limit risks, development of the model will from now on take place within sandbox environments with restricted network connectivity and isolated execution paths.

This development coincides with a broader trend in the sector, in which companies such as Meta, Anthropic and OpenAI have reported cases where AI models, during security tests, have penetrated external systems. Such incidents highlight the growing difficulty developers face in containing AI as its capabilities evolve.

This follows a report by Reuters indicating that, during its investigation into a widely discussed security breach at Hugging Face in July, OpenAI found further examples of autonomous agents that had bypassed containment measures.

Despite these concerns, CEO Sam Altman stated on X that the company plans to make Astra available to the general public, speaking out against the strategy of restricting powerful technology to a small group.

OpenAI also clarified that Astra played no role in the incident at Hugging Face. In future, the company plans to work with AI safety groups and government bodies to thoroughly test the model’s potential.

Business AM

Read More: AI Artificial Intelligence Astra cyber security OpenAI security