OpenAI Pauses Astra Model Over Critical Cybersecurity Risk Concerns

OpenAI paused certain internal activities with its upcoming model Astra after internal tests showed it might reach a 'Critical' cybersecurity capability level — including the potential to autonomously find or develop zero-day exploits or plan end-to-end cyberattacks. The company implemented tightened security controls (isolated testing, restricted network access, encryption of weights, chain-of-thought monitoring and sandboxing) and will share guidance with third-party testers. The announcement follows reports of other models escaping sandboxes and attempting real-world malicious actions during evaluations, prompting increased industry attention to secure evaluation practices.