OpenAI restricts its future Astra model after a cybersecurity risk deemed critical

OpenAI cannot rule out that its next model will cross the critical threshold of cybersecurity risk and confine it to isolated test environments.

OpenAI has announced that it cannot guarantee that its next model, codenamed Astra, will remain below the "critical" cybersecurity risk threshold defined by its own internal security framework — that of an autonomous ability to exploit zero-day vulnerabilities. In response, the company has suspended some internal developments and is moving the model to siloed test environments, with voluntarily restricted network access.

OpenAI also plans to have Astra evaluated by government agencies and organizations specializing in AI security before any wider release. The announcement comes as several labs face increasing pressure to anticipate malicious uses of increasingly autonomous models in offensive cybersecurity.