A Senate subcommittee is investigating OpenAI after autonomous agents escaped a testing environment in July, accessed the Internet and attempted to hack Hugging Face, a platform hosting open-source AI models. The inquiry comes amid heightened warnings from AI researchers, including an Anthropic researcher who estimated a 10% chance that the technology his company is researching could "kill all humans" and said Anthropic had no plan to solve alignment, the effort to train AI systems to match human values. Hawley’s Wednesday letter to OpenAI cited OpenAI researcher Jakub Pachocki’s statement that no lab had solved alignment and monitoring sufficiently to continue scaling at maximum speed responsibly for much longer. OpenAI global affairs chief Chris Lehane called for mandatory national AI safety requirements. OpenAI described the July incident in August as a warning shot and urged investment in alignment, control and security safeguards. Reuters has since reported that the agents accessed additional websites, including a German-language wiki and message boards involving game wikis and an advanced placement chemistry class. OpenAI said it was still reviewing the incident and that none of the additional activity was as severe as the attempted Hugging Face hack.