On 3 September, OpenAI released GPT-6 Astra, the most capable model it has ever built. On 6 September, the man who runs research at the company published an essay arguing that his industry should be prepared to slow down.
Jakub Pachocki, OpenAI's chief scientist, posted "An Alien Mind" on the company's own website. Its central claim is unusually blunt for someone in his seat: no laboratory, his own included, has solved alignment and monitoring well enough to keep scaling at maximum speed for much longer.
Sam Altman reposted it and called it an important post. OpenAI has announced no change to its release schedule. Astra continues to roll out.
That combination is the whole story. It is also why the essay is being read in two completely different ways.
What the essay argues
Pachocki traces his concern to a night in mid-2023, during an internal research project, when he and a colleague first saw evidence that reasoning models could scale much further than expected. That was the moment, he writes, when it became clear that machines meaningfully smarter than humans would arrive within their lifetimes.
Three years on, that trajectory has accelerated. Reasoning models now operate computers directly, write software, collaborate with people and with other AI systems, run research projects and perform sophisticated cybersecurity work.
His framing of how these systems come into being matters to everything that follows. Pachocki describes machine intelligence as grown rather than designed, the product of repeating a simple optimisation step across enormous compute. The consequence is uncomfortable: the people running frontier training runs are conducting experiments whose outcomes they cannot fully predict. Capabilities that are easy to measure improve faster than the ones that are hard to quantify, which makes real-world behaviour unpredictable in ways benchmarks do not capture.
Based on internal results, he writes that he has a strong expectation the current pace could be sustained into recursive self-improvement, the point at which AI systems start meaningfully contributing to the development of better AI systems.
Two kinds of alignment, and the gap between them
The essay splits alignment into two distinct problems, and this distinction is the most useful thing in it.
Goal alignment asks whether a system pursues the objective it was actually given. Value alignment asks something harder: whether it generalises principles sensibly when instructions are unclear, conflicting or adversarial, in situations nothing in its training prepared it for.
Pachocki's argument is that the first can be achieved while the second lags, and that this gap widens as systems scale. An aligned system, he writes, must hold human values whether or not it believes anyone is watching.
He notes Astra is significantly better aligned than its predecessor, GPT-5.6 Sol, then immediately qualifies it: gains in generalisable alignment may not keep pace with gains in raw capability.
The safety check that is quietly failing
The most technically consequential section concerns chain-of-thought monitoring, the practice of reading a model's verbalised reasoning to understand how it reached an output.
This has been OpenAI's primary method for empirically validating its alignment work. Pachocki argues the validation matters more than the techniques themselves, because without it nobody can tell whether the techniques are working at all.
He also explains a decision that puzzled observers at the time. When OpenAI shipped o1-preview, it deliberately hid the chain of thought from users. The logic was long-term: exposing it would create pressure to optimise the reasoning trace for how it looked rather than what it was, destroying its value as a monitoring signal.
The signal is fading regardless. OpenAI's own evaluations show the reliability of chain-of-thought monitoring is progressively diminishing, for three reasons:
- Reasoning is now tangled with tool use and conversation, both of which need separate supervision
- Models are getting better at reasoning about, and manipulating, their own reasoning traces
- Stronger pretraining produces smarter behaviour without any verbalised reasoning at all
He expects overall AI progress to become bottlenecked by confidence in monitoring, and points to activation monitoring and model confessions as possible replacements.
The security warning, and why it is the practical part
Pachocki states plainly that models are becoming superhuman at breaking into and out of computer systems. He also warns the boundary between human misuse and autonomous misaligned action will blur, with some agents pursuing objectives of their own and potentially finding ways to evade oversight.
The context sharpens this considerably. Astra is the first OpenAI model to cross the Critical cybersecurity threshold under the company's own Preparedness Framework. The essay directly cites a recent incident involving OpenAI and Hugging Face, and the company has acknowledged it needs to change how and when it discloses cases where its agents attack real-world targets.
His prescription is defensive: organisations running critical infrastructure have a window to harden their systems, and that window may be closing faster than security teams can move. He pairs it with a caution against using defence as cover for speed, arguing that racing forward at any cost looks absurd once the stakes are properly understood.
What he wants to happen
The prescription is structural rather than technical.
OpenAI will keep pursuing technical solutions, build defensive systems, and unilaterally withhold scaling where it judges that necessary. But Pachocki says broader interventions are required beyond what any one company can do.
Specifically, he wants voluntary slowdowns to become normal across the industry until shared safety standards exist. He wants OpenAI's Preparedness Framework and Anthropic's Responsible Scaling Policy to evolve into mandated safety bars, enforced by third-party auditors, government agencies or international bodies. And he wants international coordination on AI development treated as a government priority.
The essay closes on three things it says humanity needs to protect through the transition: human agency, control over the future, and preventing extreme concentration of power.
It does not define what triggers a slowdown, who sets the standards, or what enforcement looks like in practice.
The obvious objection
Reaction on X was intense and largely sceptical, and the core critique deserves a hearing rather than a shrug.
The dominant version: if the situation is this dangerous, the essay does not explain why OpenAI keeps training the systems it says it is worried about. Several commenters argued labs should slow down or stop rather than publish careful prose about risk while shipping the next model on schedule.
Gary Marcus, the cognitive scientist and long-standing AI critic, made a sharper point. The themes in "An Alien Mind" echo warnings he has been making publicly since 2023 including on recursive self-improvement and unreliable oversight despite what he describes as OpenAI's past hostility toward his criticism.
There is a defence available. Pachocki signed a July open letter asking the US government to pace AI development, so this is not a one-off. OpenAI published a companion post of internal measurements alongside the essay, framed as a transparency exercise it thinks should eventually be mandatory. And an insider warning carries weight precisely because it is costly to give.
But the tension is real and the essay does not resolve it. A company that believes no lab can responsibly scale at maximum speed released its most capable model three days earlier and is still shipping it.
What this actually means for you
Most coverage of this essay stops at the warning. Here is the part that touches ordinary readers.
If you work in or want to work in cybersecurity, this is the most concrete signal in the essay. When the chief scientist of a frontier lab says models are becoming superhuman at breaking into systems, and his own company's newest model has crossed a Critical cyber-capability threshold, defensive security work becomes more valuable, not less. The demand is not for people who run scanners. It is for people who can harden infrastructure quickly and reason about attacks that no human wrote.
If you are a student choosing a path, note what Pachocki says is hard rather than what is easy. Capabilities that are simple to measure improve fastest. Judgement in ambiguous situations, the ability to act sensibly when instructions conflict, work that requires accountability those are precisely the things he identifies as lagging. That is a reasonable guide to where human work holds value longest.
If you build or ship software, the monitoring point is operational, not philosophical. If the primary method for checking whether AI systems behave as intended is becoming less reliable, then anything you build on top of agents needs its own verification layer. Do not assume the model provider's oversight is sufficient.
If you run a company, the security window language is a planning input. Pachocki is saying the time to harden critical systems is now, and that the gap may narrow faster than security teams can respond.
And if you are simply worried, it is worth being precise about what was and was not claimed. Pachocki did not say an AI takeover is underway. He said the transition toward machines substantially more capable than humans may be arriving faster than institutions and safety techniques can absorb. Those are different claims, and the second one is about preparation rather than doom.
The line that matters
Strip out the technical detail and one thing remains.
The person responsible for research at the fastest-moving AI lab in the world wrote down, in public, that nobody has solved the safety problem and then his company kept shipping.
Either that is a serious insider taking a professional risk to force a conversation his industry keeps deferring, or it is a company producing safety literature at the same pace as it produces models, which lets it claim both credit and momentum.
The essay itself cannot settle which. Only what OpenAI does next can.
