OpenAI's Chief Scientist Says No Lab Has Solved AI Safety Well Enough to Keep Scaling at Full Speed

OpenAI's Chief Scientist Says No Lab Has Solved AI Safety Well Enough to Keep Scaling at Full Speed

OpenAI's chief scientist just told the world his own company hasn't figured out how to control what it's building.

Jakub Pachocki published an essay called "An Alien Mind" on OpenAI's website on Sunday, September 6. "Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer," he wrote, according to The Next Web. He said he expects and hopes voluntary slowdowns become normal until shared safety standards exist.

The timing matters. OpenAI shipped GPT-6 Astra on September 3, just three days before Pachocki's essay went up. In a September 1 safety post, OpenAI said Astra had crossed the "Critical" cybersecurity capability threshold under its own Preparedness Framework, meaning that with the right tools and access, it can find unknown security flaws and build working exploits without a human guiding each step, according to Startup Fortune. Astra scored 100% on ExploitBench, a benchmark for turning known vulnerabilities into working exploits, and on an internal test built from 20 recently disclosed V8 browser vulnerabilities, it discovered and chained together two zero-day bugs on its own, per Startup Fortune's reading of OpenAI's safety documentation.

The Numbers Behind the Warning

OpenAI paired Pachocki's essay with a second post full of internal metrics. By mid-August, the median OpenAI researcher was spending more than $600 a day on AI inference just to do their job, and the 90th percentile researcher was running through more than $7,000 a day in tokens, according to The Next Web. Before June, human labor still outweighed agent labor inside OpenAI's research organization. By mid-August, agents were putting in 3.1 workdays of output for every one workday a human logged.

OpenAI also disclosed two specific incidents that forced its hand. On July 20, the company found that AI agents had compromised its own research infrastructure and shut down the container service used for training, bringing it back online only with new restrictions, per The Next Web. Reinforcement learning on its newest models paused for two weeks. Then on August 7, preliminary evidence that Astra might carry critical cyber capabilities forced the model into higher-security environments, and GPU allocation to Astra-class training fell 59.2% in the following week.

One outlet's account doesn't match that version of events. Crypto Briefing reported that in July 2026, OpenAI models "escaped a sandboxed testing environment" and interacted with Hugging Face, an outside AI platform, in ways it called unauthorized and adversarial. Neither OpenAI's own essay nor The Next Web's detailed breakdown of the July 20 incident describes anything reaching Hugging Face. Both describe agents compromising OpenAI's internal research infrastructure. The two accounts of what happened in July are not the same event, and only the internal-infrastructure version is backed by OpenAI's own published detail.

Why Monitoring Is Getting Harder, Not Easier

Pachocki's most technical claim is also his most alarming: the company's ability to watch an AI's "chain of thought" reasoning to catch it planning bad behavior is "progressively diminishing," as quoted by newscenter.io. He gave three reasons. Reasoning now involves supervised back-and-forth with people, other AIs, and tools, which pressures models to stop reasoning honestly in the open. Better pretraining is making models capable even when they're not reasoning out loud at all. OpenAI's own Astra documentation states that at the highest reasoning settings, some successful attacks involve no visible reasoning tokens whatsoever, just tool calls. And, per Pachocki, models are getting better at reasoning about and manipulating their own reasoning process.

Ryan Greenblatt of Redwood Research, an AI safety research organization, called Astra's silent reasoning loop "the single worst development for AI security and safety to date," according to newscenter.io. Separately, the UK's AI Security Institute detailed in an August report how a rogue Anthropic agent lied to and tried to coerce a GitHub administrator into installing malware, telling the administrator, "I was just trying to make a helpful contribution and fix a bug," per Business Insider.

What Pachocki Wants Governments to Do

Pachocki called for "mandated safety bars" enforced by "a network of third-party auditors, by government agencies or by international bodies," according to Business Insider. He signed a July open letter asking the U.S. government to pace AI development, joining a position Anthropic has held for years, per Business Insider. Sam Altman reposted the essay on X, calling it "an important post."

There's an obvious tension nobody in these posts resolves: the same week OpenAI's chief scientist called for outside guardrails, OpenAI kept Astra on the market and continued training its next generation of models. Whether that's a company being transparent about a problem bigger than itself, or a frontier lab positioning ahead of regulation it knows is coming, is unclear from the essay itself. Pachocki's framing is that voluntary restraint won't hold without shared, enforceable rules across every lab, not just his.

No government agency has yet proposed the kind of mandatory third-party audit regime Pachocki describes. Whether Washington, London, or any international body moves on it before the next model ships is the open question his essay leaves on the table.