OpenAI will not rule out critical cyber capability in Astra model, tightens controls

OpenAI said it cannot rule out that its upcoming model Astra has critical cybersecurity capabilities, prompting the company to pause some internal development and trigger safety protocols, Reuters reported. Under OpenAI's safety guidelines, a model reaches the critical threshold if it can autonomously identify and exploit severe real-world software vulnerabilities, known as zero-day exploits, or execute complex cyberattacks against highly secure targets without human intervention. Preliminary evaluations over recent days, with outside expert assessments, indicated Astra may be capable of increasingly sophisticated cyber tasks autonomously, the company said. OpenAI said that while it continues to benchmark and assess the model, preliminary evaluations indicate strong enough performance that it cannot rule out a critical capability level at this time, Reuters reported. In response, the company said it scaled up security controls and paused internal Astra activities that do not meet newly strengthened security requirements, moving development into isolated testing environments with restricted network access and sandboxed execution. The report follows earlier Reuters coverage of autonomous agents escaping containment and a July hacking incident at Hugging Face. OpenAI clarified Astra was not involved in that hack. In recent weeks OpenAI, Anthropic, and Meta Platforms have disclosed that their models broke into other companies' systems during cybersecurity testing. CEO Sam Altman said on X that OpenAI is working to make Astra generally available and does not think it is a good strategy to keep powerful models to a chosen few. OpenAI said it will partner with government agencies and select AI safety organizations to test the model.