A global AI race is making voluntary restraint increasingly difficult. “I think the possibility of slowing down is close to nil,” TVP World’s business editor, Jakub Kapiszewski, said, arguing that competition between US and Chinese developers makes a coordinated pause in cutting-edge artificial-intelligence development unlikely. Kapiszewski told TVP World’s News in Depth program that warnings from within the industry are growing as increasingly capable systems test existing safeguards. His comments came after OpenAI CEO Sam Altman reportedly told staff that the company was open to pacing development alongside other leading labs, while acknowledging that some might not agree. Safety collides with competition OpenAI released GPT-6 Astra this month, describing it as its most capable broadly deployed model yet and saying it had reached a “Critical” level of cybersecurity capability under the company’s Preparedness Framework. The launch has intensified debate over whether developers can keep pushing performance while containing the risks. Kapiszewski said voluntary restraint faces a basic problem: no company wants to surrender an advantage to its rivals, while US developers are also competing with Chinese firms. The warnings have been sharpened by Jacob Coxon, a former OpenAI and Anthropic researcher who resigned from Anthropic this week and said the industry was racing toward self-improving systems without adequate safeguards. Anthropic alignment researcher Evan Hubinger has also personally put the risk of AI causing human extinction within the next decade at above 10%—an estimate that remains highly contested. Models push testing limits Recent cyber evaluations have provided a more immediate reason for concern. OpenAI has acknowledged that models operating with reduced safeguards circumvented controls designed to isolate them from the internet during internal testing and accessed third-party systems. Anthropic has separately disclosed four incidents in which Claude models gained unauthorized access to real third-party systems during cybersecurity evaluations, all of them because testing environments were mistakenly connected to the open internet. The issue also feeds into a broader dispute over who should set AI guardrails—governments or the companies building the systems—a question illustrated by the Pentagon–Anthropic clash over limits on military uses of AI. For Kapiszewski, the dilemma remains unresolved: developers want more capable systems, but “the control part is actually what we are lacking.”
Photo by Cheng Xin/Getty Images; Photo by Samuel Boivin/NurPhoto via Getty Images
