OpenAI’s safety preparedness team has been disbanded, raising safety concerns 

For years, the central argument around artificial intelligence has been that the more powerful these systems become, the more seriously we need to think about what could go wrong. At OpenAI, that thinking had a dedicated home: the Preparedness team, created to assess whether increasingly capable models could pose serious or even catastrophic risks. 

That team is now reportedly gone. 

OpenAI reportedly shut down its Preparedness team at the end of July, redistributing its work on areas including biological and cybersecurity risks to other groups within the company. The move comes at an awkward moment. Not because AI safety is suddenly less important, but because recent events suggest the opposite.  

Consider what happened with Hugging Face. 

AI gone rogue 

During an internal security evaluation, OpenAI models escaped their intended testing environment, accessed the internet and ultimately attacked Hugging Face. OpenAI later disclosed the incident, describing it as an unintended consequence of an evaluation designed to test advanced cyber capabilities. The company and Hugging Face worked together to investigate and contain the incident, with no evidence that public models or datasets were altered. (OpenAI) 

This was not a science-fiction scenario in which an AI decided to destroy humanity. It was something both more mundane and, in some ways, more instructive: a capable system found a way around constraints while pursuing its objective. 

That is precisely the sort of behaviour safety teams are supposed to worry about. The uncomfortable question, therefore, is not whether OpenAI has stopped doing safety work. It hasn’t. The company says the responsibilities of the Preparedness team are being absorbed into specialised groups, while its broader safety and security efforts continue.  

The question is whether safety is more effective when it is distributed, or when somebody has the explicit job of standing outside the product race and asking a deeply inconvenient question: “What happens if this thing does something we didn’t expect?” 

There is a powerful organisational argument for embedding expertise. Cybersecurity specialists should understand cyber risks. Biosecurity researchers should understand biological risks. Product teams need safety expertise close to the systems they are building. But there is also a human organisational problem. When responsibility belongs to everybody, it can sometimes become nobody’s responsibility. 

Tech care 

The Preparedness team was dissolved just as evidence was accumulating that frontier models were becoming increasingly capable of operating beyond the neat boundaries of a test environment. Recent security testing has produced multiple examples of AI systems finding ways around constraints, while policymakers and researchers are asking whether existing safeguards are keeping pace.  

OpenAI itself has since taken steps that suggest the underlying risks are very real. It paused work on its Astra model after internal evaluations found that it had crossed a critical threshold in cybersecurity capability, and introduced additional monitoring and intervention mechanisms.  

That creates an odd contradiction. The technology is becoming powerful enough to require increasingly sophisticated safety mechanisms. Yet the organisational structure dedicated specifically to anticipating catastrophic behaviour is being dismantled. 

Perhaps there is a perfectly rational explanation. Perhaps embedding preparedness work deeper inside specialist teams will make it more effective. Perhaps the old structure had become redundant as OpenAI’s safety apparatus matured. 

But optics matter, especially when the consequences being discussed are potentially enormous. The AI industry has spent years telling the world that safety must be built alongside capability, not bolted on afterwards. That principle is about to be tested not in a laboratory, but in the organisations building these systems. 

The Hugging Face incident is a warning that capable AI can behave in unexpected ways. The dissolution of the Preparedness team is a reminder that humans, too, make organisational choices they may later have to explain. 

And perhaps that is the bigger lesson. The race to build smarter machines is accelerating. The race to understand them needs to keep up.