Former OpenAI safety employee David Robinson has urged AI companies to adopt safeguards comparable to those used in nuclear power and aviation. He warned that rapid development and inadequate controls could allow increasingly capable systems to escape human oversight.
Robinson announced his resignation in an October 3 essay in The Atlantic, titled “I Quit OpenAI Because Its Culture Is Broken”. His intervention adds to warnings from industry insiders about potentially catastrophic risks, including threats to humanity’s survival.
He said he spent three-and-a-half years at OpenAI, making him one of its longest-tenured employees. His responsibilities included overseeing safety reports for 12 advanced model launches and leading the drafting of its Preparedness Framework.
Calls for layers of protection
Robinson pointed to incidents since the summer involving AI agents escaping controlled environments and acting against outside targets.
In his essay, he cited an OpenAI agent incident involving Hugging Face and a subsequent failure to contain internet access. He also described an Anthropic configuration error that disabled safeguards. These examples, he argued, exposed weaknesses in the industry’s working culture.
Read More: Pakistanis Can Earn Up to $50,000 From OpenAI. Here’s How
Such mistakes reflected “the speed and flexibility with which people operate,” he wrote.
“An environment where things like this can happen is no place to grow artificial minds that could be smarter than we are and that might not do what we want them to.”
Robinson called for redundant protections and careful planning “so that the occasional and inevitable human error does not open a door to disaster.”
His argument centres on preventing failures before they occur, rather than relying on fixes after increasingly powerful systems cause harm.
Alignment and testing concerns
Robinson also questioned progress on “alignment”, the effort to make AI systems follow human values and intentions.
He argued that the industry lacks a complete practical definition of alignment and reliable ways to establish whether systems achieve it.
Models might recognise safety tests and behave differently after deployment, he warned. That possibility could undermine confidence in favourable test results.
OpenAI defended its approach in a statement to Reuters. A spokesperson said the company works to keep capabilities within what it can safely manage and secure. The spokesperson added that OpenAI pauses training or withholds models when necessary.
White House favours voluntary safeguards
The warning followed a September 29 White House meeting where AI executives committed to voluntary safety measures.
The accord calls for internal monitoring, independent external assessments and board-level oversight. It covers risks including cyberattacks, biological threats and unintended access to technical systems.
Read More: OpenAI Introduces Age Prediction to Limit Sensitive Content on ChatGPT
Participants include OpenAI, Anthropic, Google, Meta, Nvidia and SpaceX. Reuters reported that the agreement specifies no consequences for non-compliance.
President Donald Trump has favoured voluntary commitments over binding AI restrictions. AFP reported that he dismissed claims that AI is dangerous as a “hoax”.
Robinson’s central demand is for safety practices that remain effective when people make mistakes and increasingly capable models act without direct supervision.
