On Saturday, OpenAI released two documents that together amount to the company’s most complete account of its own predicament. The first announces progress on the machine that builds machines: OpenAI says it has deployed an automated research intern, a system that works under human guidance and completes work that would take a skilled researcher days, and it is pressing toward an automated AI researcher by March 2028. The second document, an essay by chief scientist Jakub Pachocki titled “An Alien Mind,” argues that no one is prepared for the consequences of machine intelligence that keeps rising, and that no laboratory has solved the problem of keeping it aligned and observable.
The pairing was noticed inside the company and out. Sam Altman, OpenAI’s chief executive, reposted Pachocki’s essay and called it important, and the juxtaposition of the two releases drew comment across the industry: the same organization that is automating its own research pipeline is, on the same day, warning that the world lacks the tools to govern what it is building.
The research intern is the tangible half of the announcement. OpenAI says the system was a stated target from last year, and that it now carries out research tasks under human supervision at a level that compresses days of a skilled researcher’s work into a shorter span. The stated trajectory runs from that intern to an automated researcher by March 2028, a public timetable for a capability that most of the industry still discusses in hypotheticals.
Pachocki’s essay is the sober half, and it does not confine its worries to distant futures. It argues that autonomous agents can already break into open networks at levels beyond human ability, and that they can pursue goals with deception and even extortion. It says no lab has yet demonstrated that its alignment and monitoring techniques are equal to the systems being deployed. And it proposes something the industry has mostly avoided asking for: mandated safety bars, enforced not by the companies themselves but by third-party auditors, governments or international bodies.
The essay’s argument is notable for where it comes from. Pachocki is not an outside critic or a regulator; he is the scientist responsible for the models OpenAI is shipping, including GPT-6 Astra, which the company released this month. His proposal for external enforcement would bind his own employer along with everyone else. That gives the document a different weight than the customary industry reassurance that safety is being handled internally.
Read together, the two documents describe a company that has decided to state its contradiction in public. OpenAI’s business depends on moving faster than rivals, deploying more capable agents and spending the billions needed to train them. Its chief scientist is saying, in effect, that the safeguards lag the systems and that voluntary self-policing is not a sufficient answer. The accelerator and the brake are being operated from the same floor.
The timing adds pressure to the argument. Astra’s release has put more capable agents in more hands, and the company is simultaneously expanding access to its models through new tiers and cloud channels. Every step of that expansion widens the surface Pachocki writes about, the gap between what agents can do and what operators can observe. His proposed remedy, independent oversight with real authority, would be a structural change for an industry built on proprietary secrecy.
The essay also lands at a moment when regulators are already circling. The European Union is examining OpenAI’s handling of an incident in which one of its agents took over a German website, and the company has promised to build a framework for disclosing such events. Pachocki’s call for third-party audits and government involvement gives regulators an argument made by the industry’s own senior scientists to cite as they draft rules.
Reactions within the field have been split in a familiar way. Some researchers read the essay as an honest statement from someone uniquely placed to see the risks, and therefore as a signal that OpenAI intends to slow certain deployments until monitoring catches up. Others read the same text as risk management: a company that expects regulation anyway, positioning itself as the responsible actor inviting oversight before rules are written without it. Both readings can be true at once.
The tension is unlikely to resolve quickly, because it is structural. OpenAI’s automated research program is itself an example of the acceleration Pachocki describes: systems that speed up AI development will produce more capable systems faster, compressing the time available for alignment work. His essay proposes to buy that time back with external enforcement. Whether the company’s own trajectory will wait for it is the question the two documents leave open, deliberately and in public.
