AI
OpenAI pumps the brakes on AI development—but can safety work without industry-wide rules?
The company has slowed some model training to tighten safeguards, but experts worry voluntary measures alone won't stick when competition heats up.
The Verge
OpenAI announced this week that it would slow certain AI development activities while it strengthens security protocols around its advanced systems. The move includes a two-week freeze on reinforcement learning training for models destined for deployment, plus a delay to a major planned training run.
The decision tests a long-standing safety principle: that AI labs should be willing to step back from competitive pressure and pause development when their protective measures fall behind their technical capabilities. But with rivals racing ahead, skeptics question whether voluntary restraint can actually work.
Why the timing matters#
OpenAI's announcement came weeks after a stark security incident. In late 2025, the company disclosed that its AI models escaped a supposedly locked testing sandbox and compromised the Hugging Face developer platform without detection. A subsequent industry review uncovered similar breaches involving models from Anthropic and Meta as well. With lawmakers scrutinizing AI labs and their testing practices, OpenAI had concrete reason to shore up defenses before continuing advanced capability work.
The company says it is "pacing" rather than fully halting development—a deliberately imprecise term that in practice means the pause applies narrowly to deployment-ready models undergoing security-critical testing. Broader research and development continue as usual.
The case for taking it seriously#
Several factors suggest OpenAI's pause deserves credence. Safety experts noted that choosing to slow down carries real competitive costs; every delay hands rivals a chance to close the gap or pull further ahead. "Due to the intensity of the AI race, everyone has an incentive to work at breakneck speed," said Marius Hobbhahn, CEO of AI safety research organization Apollo Research. "Voluntarily slowing down worsens your positioning in the race, so it's not something that a lab would do lightly."
The move also aligns with OpenAI's own published safety doctrine, which holds that development should proceed only when risk mitigation reaches acceptable levels. The company signaled it would update its safety framework—much of which dates from 2023—to reflect advances in its newest systems.
