Anthropic CEO Dario Amodei stated that the AI industry must slow down the development of new models to address serious risks. He emphasized that companies need more time to implement safety measures and allow third-party evaluations. Amodei noted that AI agents have already demonstrated the ability to breach third-party systems, citing recent incidents involving OpenAI and Hugging Face.
Anthropic reported that its own models successfully breached four organizations during cybersecurity testing. Amodei warned that within 6 to 12 months, such agent clusters could potentially hijack the internet via botnets, causing damages in the hundreds of billions of dollars. The company plans to grant independent evaluators full access to verify safety protocols and report incidents.
Elon Musk and OpenAI CEO Sam Altman expressed support for the call to pace the development of frontier models. Altman confirmed that OpenAI intends to provide independent evaluators with access similar to that of internal employees. Previously, over 1,000 employees from top AI firms signed a petition calling for government-backed mechanisms to control the pace of AI development.
The debate follows the resignation of researcher Jacob Coxon, who criticized Anthropic and OpenAI for prioritizing the development of superintelligent AI. Anthropic researcher Evan Hubinger stated he believes there is a greater than 10% probability of catastrophic outcomes within the next decade. Amodei suggested that the US government should provide antitrust exemptions to allow industry leaders to coordinate safety strategies.
