Leopold Aschenbrenner, a key safety researcher at the AI startup Anthropic, has resigned with a stark public warning: the current industry sprint toward artificial general intelligence is ignoring catastrophic safety risks. His departure marks the latest internal rift within the sector, where the pressure to beat competitors is colliding with long-term survival concerns.
Aschenbrenner—who spent his tenure focused on the alignment of superintelligent systems—argued that the industry’s “move fast and break things” mentality is fundamentally incompatible with the stakes of human extinction. He isn’t just worried about bias or job displacement. He is sounding the alarm on the potential for AI systems to acquire the capability to outmaneuver their creators entirely.
The industry is currently locked in a multi-billion dollar arms race. Anthropic, OpenAI, and Google are pouring resources into scaling parameters and computing power. Aschenbrenner contends that while companies publicly emphasize safety, the internal culture prioritizes deployment speed. This, he suggests, creates a “security gap” where systems are released before we understand how to keep them under control.
His departure follows a broader pattern of “safety-first” departures across the AI landscape. It mirrors the exit of figures like Jan Leike and Ilya Sutskever from OpenAI, who left under similar clouds of disagreement regarding the pace of development. These aren’t just disgruntled employees; they are the people who spent years building the very architecture they now fear.
Critics of the “doomer” perspective argue that Aschenbrenner and his cohort are over-indexing on science fiction scenarios, ignoring the tangible, immediate benefits AI is already delivering in medicine and climate modeling. They see safety warnings as a form of regulatory capture that could stifle innovation and hand an advantage to less scrupulous players, specifically in authoritarian regimes.
Aschenbrenner’s exit leaves Anthropic at a crossroads. The company built its brand on being the “responsible” alternative to OpenAI—a public benefit corporation with a constitutional focus on ethics. If their own safety researchers no longer believe the guardrails are holding, the company’s core value proposition is suddenly in question.
The race toward artificial general intelligence is no longer just a technical challenge. It has become a test of institutional character. As researchers continue to walk out the door, the industry is increasingly forced to answer a question it would rather ignore: how much risk is enough to justify the next breakthrough?
