David Robinson spent three and a half years inside OpenAI’s safety operation, working on transparency, preparedness and the public-facing reports meant to explain how the company’s models behave. By early September, he was already signalling unease: OpenAI was changing rapidly, he wrote, but he did not know whether it was changing fast enough.¹

That anxiety grew against a rougher backdrop. OpenAI’s safety ranks had been unsettled by earlier departures, including safety head Johannes Heidecke, while the company faced scrutiny after security incidents and reports of model misbehaviour. It also said it had parted ways with three researchers over alleged violations of policies governing sensitive information.¹

Last week, Robinson resigned from the Safety Systems team. In an essay published Saturday, he argued that the pace of the work made internal reform far harder than it sounded. “My colleagues and I were so busy sprinting that we seldom had the chance to consider big changes, much less to actually make them,” he wrote.²

His conclusion was pointed: “Stronger incentives for safety — coming from outside the company — are a big part of getting this right.” Robinson said he now hopes to help the public understand the risks he encountered and to increase pressure on OpenAI and rival firms to act more cautiously.²

OpenAI rejects the implication that capability is being allowed to outrun control. A spokesperson said the company is strengthening safety and security practices as its models improve, using third-party evaluators and earlier monitoring for concerning behaviour. The company’s central promise is that it will not push ahead regardless of risk: “We pause training or hold back models when we need to slow down.”²

The divide is less about whether advanced AI needs guardrails than about who can reliably enforce them. Robinson has chosen external leverage; OpenAI says its internal systems can still do the job.