Senior warning on AI risks

Sam Altman, the chief executive of OpenAI, used unusually blunt language this week to describe what he thinks may be the first truly catastrophic AI misuse event. In a new interview clip circulating on X, he said rapid progress in AI makes a “world-shaking” cyberattack feel plausibly near-term, even possibly this year. The warning landed the same day OpenAI published a new policy blueprint arguing that the country needs a new social contract for the age of “superintelligence,” not just better models and better guardrails. (latestly.com) That pairing matters. Altman was not only talking about technical failure. He was tying frontier AI to public safety, labor disruption, wealth concentration, and democratic legitimacy all at once. OpenAI’s new paper says outright that incremental policy updates will not be enough as systems become more capable. It frames the transition as large enough to strain the institutions that distribute income, regulate markets, and absorb shocks. (openai.com) The cyberattack warning did not come out of nowhere. OpenAI’s own Preparedness Framework already treats cybersecurity as one of the core “tracked” risk categories for frontier models, alongside biological and chemical capabilities and AI self-improvement. The framework is built around the idea of “severe harm,” including scenarios that could cause thousands of deaths or hundreds of billions of dollars in damage. In other words, the company has already formalized the possibility that advanced models could help produce national-scale disasters. Altman just translated that internal safety language into plain English. (openai.com) He also narrowed the timeline. That is the striking part. AI executives have spent years talking about long-run existential risk. Altman’s recent public warnings have shifted toward immediate misuse. In July 2025, he warned of an impending AI-enabled fraud crisis, especially in systems that still rely on voice authentication. Now the emphasis has moved from scams and impersonation to the possibility of a cyber event big enough to shake the world. The pattern is clear. The nearer AI gets to real operational competence, the less the risk looks like science fiction and the more it looks like infrastructure security. (axios.com) That helps explain why OpenAI’s new document is not mainly a safety memo. It is an industrial policy document. The paper argues that if AI drives huge productivity gains, the benefits cannot simply flow to the owners of models, chips, and data centers. It floats taxes tied to automation, a public wealth fund that would give citizens a stake in AI-driven growth, pilots for a 32-hour workweek, and even a “right to AI” that treats access to advanced systems more like a utility than a luxury product. These are not side notes. They are OpenAI’s admission that deployment itself is becoming a governance problem. (latestly.com) There is an obvious tension here. OpenAI is racing to build more capable systems while publishing papers about how society should absorb the blast radius if that race succeeds. The company’s new page says these ideas are intentionally exploratory and meant to start a democratic process, not end one. It is also putting money behind that claim, offering policy fellowships, research grants, and API credits, and opening a workshop space in Washington in May. That is a concrete sign that the argument has moved out of the lab and into the capital. (openai.com)