On Monday, OpenAI unveiled a framework designed to assess and address the “catastrophic risks” associated with the powerful artificial intelligence (AI) models it develops. The framework, presented in a 27-page document, outlines OpenAI’s commitment to monitoring and sharing warnings about potential risks linked to the misuse of its AI models, such as the creation of harmful weapons, malware dissemination, or social engineering attacks.
OpenAI’s preparedness team, established in October, will employ a matrix approach to evaluate four risk categories—cybersecurity, persuasion, model autonomy, and chemical, biological, radiological, and nuclear threats. Models with higher risk scores will not be deployed, and mitigation measures will be implemented for those with medium or lower risk scores. OpenAI’s framework is led by its CEO, with oversight from the company’s board, ensuring governance and risk management. The preparedness team, under the leadership of MIT AI professor Aleksander Mądry, focuses on addressing potential threats from AI technologies.
OpenAI’s broader safety strategy includes efforts to mitigate biases, hallucinations, and misuse, aligning with a voluntary commitment made to the Biden administration alongside other technology companies. The company has also joined forces with major tech firms to establish an industry watchdog group aimed at regulating AI development. This comprehensive approach reflects OpenAI’s dedication to building a safe and trustworthy AI ecosystem while addressing evolving risks associated with AI advancements.
