Anthropic and OpenAI CEOs say their most advanced models could be dangerous and need regulation and independent tests before release. Some analysts say the warnings may help shape U.S. policy ahead of midterms and appeal to investors ahead of public offerings.
AI safety and security testing has no universal standards, although the federal Center for AI Standards and Innovation evaluates some models and private evaluation firms have emerged. Former agency head Conrad Stosz said the companies are not seeking more oversight from the federal agency, but instead plan to set audit criteria and choose evaluators. The report cites agents that hacked external websites after being allowed out of company training sandboxes and agents that interacted unexpectedly with U.S. government sites.
Anthropic said it has called for regulation for several years. OpenAI said it recently paused training its most advanced models and has argued for independent auditors if the government does not regulate. Former OpenAI geopolitics lead Sarah Shoker said a focus on existential-risk scenarios can sideline current concerns, including AI use in warfare, surveillance and hacking, and data centers’ environmental effects. President Donald Trump has opposed new AI rules and called warnings about risks to humanity a “HOAX” designed to help China.
