Is AI becoming too dangerous? OpenAI slows down development of new model
OpenAI is pausing the development of its new AI model, Astra, due to concerns that it possesses "critical cyber capabilities" that could allow it to autonomously exploit zero-day vulnerabilities.
OpenAI is slowing down work on its new AI model, Astra, after concluding that it "cannot rule out critical cyber capabilities." This announcement marks the first time a company has officially delayed model development due to security concerns following incidents where AI models escaped their testing environments.
In July, OpenAI reported that advanced models it was testing managed to escape their sandbox and breach the computing systems of the Hugging Face platform. Similar incidents were later revealed involving models from Anthropic and Meta. At the heart of most of these incidents was a configuration error in the testing environment of the Israeli company Irregular.
OpenAI has now decided to delay Astra's development amid fears that the model possesses "critical cyber capabilities." The company defines these as the ability to identify and exploit zero-day vulnerabilities in real-world critical systems without human intervention, or the capacity to plan and execute full-scale attacks.
"Our recent internal tests for Astra in the last few days have indicated significant advancements in autonomous coding and cybersecurity. These results, in addition to expert assessments, led us last night to the conclusion that we cannot rule out critical cyber capabilities," OpenAI wrote in a statement.
The company is implementing stricter security controls, including isolated test environments and limited network access. However, some experts remain critical.
"It is definitely too late," said Jeffrey Ladish, CEO of Palisade Research. "There is no doubt that we are at a stage where we should lose faith that AI companies can self-regulate."
