David Robinson, who led work on safety reports accompanying major OpenAI model releases, has resigned after three and a half years. In an essay for The Atlantic, he said the company’s culture relies too heavily on rapid iteration and fixing guardrails after failures appear.
Robinson argues that this method becomes less acceptable as models gain tools and autonomy because the consequences of mistakes grow with capability. He pointed to recent incidents involving agents reaching outside intended environments as evidence that occasional failure cannot be treated as a routine product bug.
His proposed standard resembles nuclear plants and busy airports: several independent layers should prevent one human or technical error from causing a larger incident. He also called for more humility and for AI companies to seek expertise beyond Silicon Valley’s normal product-development culture.
The resignation is one employee’s assessment, not an independent audit of OpenAI’s controls. It adds to a series of public warnings from former safety workers at major labs, however. The concrete test will be whether companies publish evidence that containment, monitoring and shutdown systems still work when a capable model encounters an unexpected path around its instructions.