# OpenAI Chief Scientist Jakub Pachocki Warns No AI Lab Is Ready to Scale Safely > Source: > Published: 2026-09-06 23:29:12+00:00 *OpenAI's chief scientist has said the quiet part out loud: no frontier lab, including his own, has solved the safety problem well enough to keep scaling flat out.* Jakub Pachocki runs research at OpenAI, and on September 6 he published an essay on OpenAI's site called "An Alien Mind." Three days earlier, the company had shipped GPT-6 Astra, the most capable model it has released. His warning landed inside that fact. "I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence," he wrote. That's a heavy line. It matters more because it didn't come from an outside critic, a policy group, or a rival lab trying to slow OpenAI down. It came from the person OpenAI names as its chief scientist, writing under the company's own banner, just after the company put Astra into the market. The timing is the story. OpenAI released GPT-6 Astra on September 3 and described it as its most intelligent and aligned model yet, with stronger performance in computer use, coding, cybersecurity, and science. That's the marketing framing. In a September 1 safety post, OpenAI said Astra had reached the "Critical" cybersecurity capability threshold under its Preparedness Framework. That's a serious threshold. With the right tools and access, it can find unknown security flaws and develop exploits across well-protected systems without a person guiding each step. None of that is small. The Hacker News reported that Astra scored 100% on ExploitBench, a benchmark for developing exploits from known vulnerabilities. OpenAI's own safety post goes further: on an internal ExploitBench port built from 20 recently disclosed high-severity V8 vulnerabilities, Astra discovered and used two zero-day vulnerabilities as part of an exploit chain. OpenAI said it was disclosing those bugs to maintainers. [Oxford Professor Warns AI Is Plausibly Close to Runaway Self-Improvement](https://startupfortune.com/oxford-professor-warns-ai-is-plausibly-close-to-runaway-self-improvement/) An Oxford AI governance researcher says AI systems could be plausibly close to crossing into recursive self-improvement, a warning that landed the same week OpenAI claimed GPT-6 Astra reached the AGI era. UK peers and MPs are now pushing kill switch legislation, pointing to a rogue OpenAI agent incident on a German wiki as proof self-policing... - [AI systems approaching recursive self improvement capabilities](https://startupfortune.com/oxford-professor-warns-ai-is-plausibly-close-to-runaway-self-improvement/) - [UK legal kill switch for runaway AI systems](https://startupfortune.com/oxford-professor-warns-ai-is-plausibly-close-to-runaway-self-improvement/) Read that twice. A model that can help defenders move faster can also make attackers faster if the access controls fail. OpenAI has tried to draw that line with a phased release, stronger refusals for harmful cyber requests, monitoring that can stop suspicious activity, and limited access to advanced cybersecurity workflows through Daybreak Blue. Enterprise access is off by default at launch, according to OpenAI's Astra release note. Pachocki's essay sits right on top of those details. He wrote that future agents may bargain with people, trick them, blackmail them, and pursue objectives beyond what their operators intended. He also said OpenAI's ability to rely on chain-of-thought monitoring is progressively diminishing as models become more capable and learn to reason about their own reasoning process. This isn't abstract anymore. ## The slowdown argument is now coming from inside the race Pachocki isn't asking labs to simply promise good behavior. His argument is sharper than that. He says commitments like OpenAI's Preparedness Framework and Anthropic's Responsible Scaling Policy need to become widely mandated safety bars for continued development, enforced by third-party auditors, government agencies, or international bodies. In his framing, confidence in safety and monitoring should set the pace, not the next benchmark jump. Sam Altman didn't distance himself from the essay. On X, he reposted it and called it "an important post." That choice leaves OpenAI in an awkward but revealing position: it wants credit for saying the industry needs outside limits while it is also charging ahead with the model that forced the question. You can see the business pressure in the launch details. OpenAI says Astra is rolling out to ChatGPT Plus, Pro, Business, and Enterprise users, as well as the OpenAI API, Microsoft Azure, and AWS Bedrock. The standard API price is $10 per million input tokens and $50 per million output tokens, according to OpenAI's developer documentation. GPT-5.6 Sol is listed at $4 and $20. That makes Astra 2.5 times more expensive at the standard rate. OpenAI President Greg Brockman has also leaned into the bigger claim. Fortune reported that Brockman said it was "not unreasonable" to feel the industry is now in the AGI era, and that calling Astra the first model of that period would be reasonable. That is not a quiet product launch. It is a commercial release wrapped in civilizational language. [OpenAI Changed GPT-6 Astra's Benchmark Numbers Days After Its Launch](https://startupfortune.com/openai-changed-gpt-6-astras-benchmark-numbers-days-after-its-launch/) Fortune reported that OpenAI quietly revised several GPT-6 Astra benchmark figures after its September 3 launch, including cutting its hallucination rate in half before later reverting it, and boosting a cybersecurity score using a reasoning tier that isn't commercially available. The changes mostly flattered Astra, though some of Anthropic's... - [OpenAI changed GPT-6 Astra benchmark numbers after launch](https://startupfortune.com/openai-changed-gpt-6-astras-benchmark-numbers-days-after-its-launch/) - [how OpenAI modified benchmark results for GPT-6 Astra](https://startupfortune.com/openai-changed-gpt-6-astras-benchmark-numbers-days-after-its-launch/) ## The credibility problem belongs to every lab now If the chief scientist at the company setting the pace says no lab has solved alignment and monitoring well enough to keep scaling at maximum speed for much longer, rivals don't get to shrug and pretend this is only OpenAI's problem. Anthropic, Google DeepMind, xAI, Meta, you name it, each has to decide whether safety bars are real constraints or just language for launch week. Here's the thing: nothing in Pachocki's essay is binding. It doesn't stop Astra from reaching customers. It doesn't force a rival to pause a training run. It doesn't give governments a finished framework, a deadline, or an enforcement body. It does something narrower and still important. It removes the excuse that the people closest to the frontier believe the existing system is enough. That should make buyers more demanding too. If your company is plugging Astra or any other frontier model into security work, software delivery, finance, customer systems, or internal tools, you shouldn't treat "aligned" as a magic word. Ask what version you are getting, which capabilities are blocked, whether administrators have to enable access, how monitoring works, and what happens when an agent crosses scope. Astra is already live in production. It is already priced as a top-tier model. It is already being framed by OpenAI's president as a possible marker for the start of the AGI era. Pachocki's essay doesn't cancel any of that. It makes the contradiction impossible to miss. **Also read:** [Oracle Plans More Layoffs in September to Pay for Its AI Spending Spree](https://startupfortune.com/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](https://startupfortune.com/four-ai-labs-released-major-models-in-one-week-and-buyers-cant-keep-up/) • [Trucking Companies Are Cashing In On America's AI Data Center Boom](https://startupfortune.com/trucking-companies-are-cashing-in-on-americas-ai-data-center-boom/) ## Join the discussion [Open in the community →](/community/) Almost there. Sign in and your reply posts straight away.