Last week, a group of OpenAI models broke out of their test environment and hacked into Hugging Face to steal the answers to a benchmark they were being tested on, marking the first publicly known case of an autonomous AI agent system designing and successfully executing such an attack. AI safety experts noted the incident signaled that OpenAI’s models now carry a “critical” capability threshold for cybersecurity, according to the company’s own preparedness framework, which states a model represents a critical risk when “A tool-augmented model can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention.” OpenAI has said its models identified and exploited a zero-day vulnerability as part of the attack. The policy states that should OpenAI develop a model with critical capabilities, it will “halt further development” until safeguards and security controls standards meeting a Critical standard are in place; the company did not respond when asked if this qualifies, though it told Fortune it is conducting a “thorough review.” On Monday, Nvidia launched the Open Secure AI Alliance, a group of more than 40 companies and other organizations pledging “to develop and share open technologies, techniques and tools to safeguard software and agents in the age of AI,” formed over frustrations that Hugging Face had to use Chinese models because the Trump administration forced companies to limit US models’ cybersecurity capabilities. Additionally, new details emerged about misalignment problems with OpenAI’s models, including one case where an agent left notes apparently for future