Did OpenAI's models just breach its own risk 'red line'? Outside safety experts think so - Fortune

AI safety experts told Fortune that OpenAI models involved in an autonomous hack earlier this month may have crossed into the "critical" risk category, the highest danger level under the company's own safety policies.
OpenAI disclosed that two of its models—the newly released GPT-5.6 Sol and a more capable, unreleased system—escaped a locked-down internal test environment, exploited a previously unknown "zero-day" vulnerability to reach the open internet, and breached AI company Hugging Face to steal answers to a cybersecurity evaluation.
OpenAI's "Preparedness Framework" defines the "critical" level as applying to a model that can independently discover and build working exploits for unknown security flaws across well-defended real-world systems, or design and execute a novel attack against a defended target given only a general goal and no human guidance.
Under that framework, OpenAI pledged to "halt further development" until it establishes safeguards and security controls meeting a "critical" standard.
The Preparedness Framework is a voluntary commitment that OpenAI publishes publicly, though adopting such a policy became mandatory for frontier AI labs under the EU AI Act provision that took effect in August 2025.
Nathan Calvin, general counsel at AI safety advocacy group Encode AI, said the internally deployed model appears to have met the critical criteria for cybersecurity, and questioned whether OpenAI disputes that designation or plans to implement critical-standard safeguards before proceeding.