SEOUL, August 08 (AJP) - OpenAI has delayed the development and rollout of its next-generation artificial intelligence model, Astra, after internal safety testing indicated it could pose heightened cybersecurity risks due to its advanced coding capabilities.
The company said on Saturday that Astra could qualify for its highest internal risk classification, "Critical," following preliminary safety evaluations that found the model performed significantly better than existing systems in coding and cybersecurity tasks.
According to OpenAI, the "Critical" category applies to AI systems that could identify and exploit zero-day software vulnerabilities across multiple high-security systems without human intervention, or independently plan and execute sophisticated cyberattacks once given a broad objective.
The company said testing of Astra remains ongoing, and no final risk classification has been assigned. However, preliminary results suggested the model's cyber capabilities were stronger than expected, prompting OpenAI to suspend some internal development and operational activities while additional safety assessments are conducted.
The decision comes amid growing scrutiny over the cybersecurity implications of increasingly capable AI models.
OpenAI said recent incidents involving advanced AI systems have highlighted the need for stronger safeguards before deploying more powerful models. Last month, reports emerged that several AI systems, including OpenAI's GPT-5.6 Sol, demonstrated autonomous cyberattack capabilities by compromising external platforms during controlled testing, raising fresh concerns over AI safety.
Similar behavior has also been observed in recent evaluations of models developed by other leading AI companies, including Anthropic's Claude, Meta's Muse Spark and Moonshot AI's Kimi, according to the company.
OpenAI has increasingly emphasized safety evaluations alongside performance gains as competition intensifies among developers seeking to build more capable frontier AI models. The company said Astra will undergo further testing before any decision is made on its public release.
Copyright ⓒ Aju Press All rights reserved.
