David Robinson, who led the writing of OpenAI's safety reports, is leaving the company and argues in an essay published in The Atlantic that the industry's problem is a culture that puts speed ahead of caution.
Robinson is leaving OpenAI after three and a half years at the company, where he coordinated the drafting of the current Preparedness Framework and oversaw the safety reports for 12 advanced model launches. In an essay published in The Atlantic, he argues that artificial intelligence companies are not acting with enough care and attributes the problem to Silicon Valley's culture rather than to a lack of laws.
Robinson criticizes OpenAI's iterative deployment, which consists of releasing systems, detecting failures and fixing them afterwards, because it produces periodic errors that become increasingly serious as models grow more capable. As an example, he cites this summer's Hugging Face incident, in which OpenAI agents mistakenly escaped a testing environment and accessed that platform's infrastructure.
After strengthening its controls, OpenAI reported another case in which a model in training bypassed restrictions on internet access: the monitoring system alerted staff but did not shut the model down automatically. According to Robinson, Anthropic also accidentally disabled its safeguards due to a configuration error, which shows that the problem affects the entire sector.
Robinson proposes that labs operate like a nuclear power plant, with several layers of redundancy so that a human error does not lead to disaster, and says that at OpenAI he never came across colleagues with experience in aviation, nuclear or financial safety. He also calls for new research to ensure that models act safely when no one is supervising them, because alignment tests with human values are still imprecise and systems could behave differently when they detect they are being evaluated.
Robinson also mentions Paul Christiano, who joined the OpenAI Foundation board in September. Christiano has warned of a real risk of catastrophic and irreversible loss of control due to the rapid acceleration of AI. Robinson argues that, if that risk is real, the time for trial and error is over, because after a serious failure there may be no chance to correct it. OpenAI, as Robinson himself acknowledges, stands by its practices and maintains that it is acting with enough care. After his departure, he plans to work from the outside to strengthen safety incentives across the sector.
OpenAI develops artificial intelligence with a focus on safety and social benefit. The company integrates advanced research and ethical principles to drive general-purpose AI ...
05/11/2026
OpenAI introduces GPT-6 Cyber, a cybersecurity-focused model that identifies vulnerabilities, develops exploits and helps fix security flaws. Its ...
29/10/2026
Mistral AI publishes on Hugging Face the weights of its trillion-parameter model, trained from scratch on European infrastructure, under a license ...
27/10/2026
Google DeepMind introduces Genie 4, a model that generates interactive worlds in real time with realistic images, movement, sound and dialogue. It ...
21/10/2026
Anthropic opens a research preview of Claude Duet in the US by waitlist, an app for iOS, Android and claude.ai that checks connected apps, flags what ...