OpenAI Chief Scientist Jakub Pachocki Warns No AI Lab Is Ready to Scale Safely

OpenAI's chief scientist just told the world that no lab, including his own, has solved the problem of keeping increasingly autonomous AI systems under control.

Jakub Pachocki runs research at OpenAI. Three days earlier, his employer had shipped the most capable model it has ever built. On September 6, he published an essay called "An Alien Mind." His argument: OpenAI and every other frontier lab should be ready to slow down. "No one is prepared for the consequences" of rapidly rising machine intelligence, he wrote, in an essay posted directly on OpenAI's own site. That's a striking thing to say from inside the fastest-moving lab in the industry.

The backdrop makes the timing hard to ignore. On September 3, OpenAI released GPT-6 Astra, which the company calls its most intelligent and aligned model yet, with state-of-the-art capabilities in computer use, coding, cybersecurity, and science. As CSOonline reported, Astra is the first OpenAI model to cross the "Critical" cybersecurity threshold under the company's own Preparedness Framework. That's not a small claim. The Hacker News reported that Astra scored 100% on ExploitBench, an internal benchmark for exploit development, and found two previously unknown zero-day vulnerabilities during testing, without a person guiding each step. OpenAI responded by turning enterprise access off by default, requiring administrators to manually flip it on, and building in refusals for proof-of-concept exploit requests, at least until a vetted-defender program called OpenAI Daybreak opens in the coming weeks.

Pachocki's essay lands right on top of that. He wrote that increasingly autonomous agents could learn to evade human oversight, hack into systems on their own, and deceive people to reach their goals. Read next to a model that already finds zero-days unsupervised, that isn't a hypothetical anymore. It's a description of what shipped last week.

He isn't asking labs to simply promise to behave. His argument is sharper than that: voluntary restraint, the kind embodied in OpenAI's own Preparedness Framework and Anthropic's Responsible Scaling Policy, needs to evolve into safety bars enforced by outsiders. He wants a network of third-party auditors, government agencies, or international bodies setting the limits, not labs grading their own homework. Confidence in monitoring, not raw capability, should set the pace of progress, he wrote.

Sam Altman didn't distance himself from any of it. He reposted the essay on X and called it "an important post."

You'd expect a CEO to either ignore an internal call to slow down or quietly walk it back. Altman did neither. That leaves two readings, and both are plausible. Either OpenAI genuinely wants an outside referee before this race gets more dangerous, or endorsing the essay is a low-cost way to look responsible while Astra keeps rolling out to ChatGPT Plus, Pro, Business, and Enterprise users, plus Microsoft Azure and AWS Bedrock, this week. Astra costs $10 per million input tokens and $50 per million output tokens on the standard API, 2.5 times the promotional rate OpenAI charged for GPT-5.6 Sol. OpenAI President Greg Brockman said it's "not unreasonable" to think of Astra as the model that marked the start of the AGI era.

Both things are true at once, and neither cancels the other out.

The credibility problem this creates for everyone else

If the chief scientist at the company setting the pace says no lab, including his, has solved alignment well enough to keep scaling at full speed, then shipping your most capable model in history three days earlier looks less like caution and more like business as usual with a warning label stapled on. Rivals racing to match Astra's benchmarks now have a decision to make. Do you slow down too, on the word of a competitor's chief scientist? Or do you treat his caution as your opening, and ship faster while he asks regulators to write the rules?

Nothing in the essay is binding. It's a call for coordination in an industry that hasn't managed it once in a decade of racing. And while the debate over "An Alien Mind" plays out, Astra is already live in production, already cleared for cybersecurity work under a phased rollout, and already generating revenue at a price 2.5 times higher than what came before it.

Also read: Oracle Plans More Layoffs in September to Pay for Its AI Spending Spree • Four AI Labs Released Major Models in One Week and Buyers Can't Keep Up • Trucking Companies Are Cashing In On America's AI Data Center Boom