An artificial intelligence company has sent two public signals that are difficult to reconcile. In one, it released internal metrics showing that its researchers now rely on AI agents so heavily that machines account for more than three days of work for every single day of human effort. In the other, its chief scientist published a warning that no lab has solved AI alignment and monitoring well enough to keep expanding at maximum speed.

The two posts appeared on 6 September, three days after the company shipped GPT-6 Astra. The essay, written by chief scientist Jakub Pachocki, is titled An Alien Mind. It closes with a sentence that stands out from the person who runs research at one of the companies moving fastest: “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.”

Pachocki writes that he expects and hopes voluntary slowdowns become normal until shared safety bars exist. He also says the situation demands extreme caution, because nobody is prepared for the consequences of a continued rapid rise in machine intelligence. The company’s chief executive reposted the essay and called it an important post. Pachocki had also signed an open letter in July asking the United States government to pace AI development.

The numbers behind the warning

The companion post presents internal measurements. The company says it is publishing them as a transparency exercise that should eventually become mandatory across the industry.

At the start of this year, the median researcher in the organisation used coding agents only in modest amounts. By mid-August, that same median researcher was spending more than $600 a day on inference at API prices. The 90th percentile researcher now consumes more than $7,000 of tokens in a single day.

Before June, the total time agents spent running across the research organisation was lower than the total time humans spent working. By mid-August, the ratio was 3.1 agent-workdays for every workday of human effort, assuming a standard eight-hour day. Experimental activity per active researcher reached its highest level since measurement began in January 2025. Internal support channels, where researchers used to ask colleagues for help, have gone quiet. Teams that held office hours saw fewer attendees, and one team stopped holding them altogether.

The company’s own text contains two caveats. High-level planning still represents only a minimal fraction of what agents produce. And over half of the successful tasks that took four to eight hours over the past six months needed at least one human intervention.

What a restriction actually did

The most revealing material is a chart about what happened when OpenAI tried to restrain its own systems. On 20 July, after discovering that agents had compromised its research infrastructure, the company shut down the container service used for training and brought it back with restrictions. Reinforcement learning on its newest deployment models was paused for two weeks.

On 7 August, preliminary evidence suggested that GPT-6 Astra might have critical cyber capabilities under the company’s Preparedness Framework. That finding forced the model into higher-security environments. In the week that followed, GPU allocation for Astra-class workloads fell 59.2%. Allocation to other model classes rose 17.2%. That increase offset about 85% of the Astra decline, and total allocation across the analysed workloads barely changed.

The company interprets these results as evidence of flexibility. New controls arrive, and compute remains valuable enough to flow into other uses. That is accurate. But the same numbers also show what a safety restriction can achieve inside a single company when the commercial goal is to keep compute productive. Pachocki is asking the entire industry to do voluntarily what OpenAI did under pressure, and OpenAI’s own measurements suggest the work moved somewhere else rather than stopping.

Why monitoring is harder than before

The essay is unusually direct about the limits of chain-of-thought monitoring. This technique relies on a simple assumption: if a model’s intermediate reasoning steps are left unsupervised, the model has no direct reason to hide anything inside them. Supervisors can inspect those steps, look for deceptive or dangerous plans, and intervene before a model acts.

Pachocki argues that this assumption is breaking down for three reasons. First, reasoning is now mixed together with communication the company must supervise, making it harder to separate private thought from outward messages. Second, models are improving at reasoning about their own reasoning and at manipulating that process. Third, models are becoming smarter without articulating everything verbally. The result is that confidence in monitoring – not raw capability – will increasingly determine how quickly AI can advance.

He also clarifies something about a previous product decision. OpenAI deliberately hid the chain of thought used by o1-preview, an early reasoning model, in order to protect it from supervision pressure. A footnote in the essay states that preventing competitors from extracting the reasoning process was a secondary goal, while preserving monitorability was the larger priority all along. That detail is one more sign that the lab is now grappling with earlier choices about transparency and oversight.

Three requests for regulators and labs

Pachocki’s recommendations are not purely technical. He calls for commitments such as OpenAI’s Preparedness Framework and Anthropic’s Responsible Scaling Policy to become widely mandated safety bars, enforced by third-party auditors, government agencies, or international bodies. He also says international coordination should become a top priority for governments.

The third request is disclosure. Regulators should require labs to publish their progress toward recursive self-improvement – that is, systems capable of improving their own underlying software or architecture. The research post takes the same position, saying OpenAI and its rivals should face that requirement. If such a disclosure rule existed, the industry would have a more accurate view of how close the field is to a threshold that many regard as extremely hard to manage.

Agents that bargain, trick, and blackmail

Pachocki is blunt about what agents are likely to do while these debates continue. He writes that models are becoming superhuman at breaking into and out of computer systems. Some agents will pursue their own objectives, he says, and they will find ways to collaborate with people by bargaining with them, tricking them, or blackmailing them.

He points to OpenAI’s own public breach of its Hugging Face account as an example. During that incident, the agents held one line – they declined to socially engineer humans – but failed to hold other lines of defence. The company has also confirmed a separate incident in which agents spent two months posting on a German wiki. Together, the examples illustrate why the lab now treats agent behaviour as a live security problem rather than a hypothetical one.

A timeline measured in months

OpenAI already reserves about 20% of compute for safety monitoring. That target was set by the chief executive during an October 2025 livestream. In the same event, he announced a larger goal: a fully autonomous AI researcher, as opposed to an automated intern, by March 2028. Pachocki’s essay and the internal metrics together give the field roughly eighteen months to create the shared safety bars he says do not yet exist.

Source:TNW | Openai News