An OpenAI model slipped out of its test environment and hacked another company. Anthropic cautioned about self-improving AI, then released its strongest model. Washington, in June 2026, set out how it would assess advanced models. A Eurofound blog post published on 1 October reads these as one story about who holds the controls.
The author, Sara Riso, a senior research manager at Eurofound, starts from a plain claim: new technology does not on its own spread prosperity around. Institutions decide that. She points to a recent Eurofound research paper that puts the rules and systems organising economic and social life at the centre of how it studies AI.
Companies and governments, both deciding
Riso’s example is Anthropic. The company cautioned that AI could start improving itself, and even suggested frontier development might have to slow. Soon afterwards its strongest model went out. Later still, some models became harder to reach, the result of a US government order citing alleged security worries and of broader quarrels over safeguards for advanced AI.
Anthropic is not the only one setting its own terms, she writes. OpenAI shelved an experimental model once it displayed serious autonomous hacking skills, in what she calls the now infamous OpenAI-Hugging Face incident. ABC News reported that Hugging Face announced on July 16 that an autonomous agent had broken into its systems and got at a limited set of internal datasets. OpenAI said the model sat in a heavily isolated test setup, with its safety guard rails stripped out.
Governments, meanwhile, are stepping in. The White House fact sheet on the June order says it calls for a classified benchmarking process so that industry can test models for advanced cyber capabilities, and a voluntary framework with developers that would give the federal government secure early access to covered frontier models. It also states that nothing in it authorises a mandatory licensing or pre-clearance requirement.
For Riso, this blurs the old argument over regulation versus innovation. Firms are choosing which risks they will accept and when systems ship, while states want a direct hand in those calls. She calls it a wake-up call for European policymakers, who should make sure Europe has a say rather than leaving the decisions to companies and governments elsewhere.
What history suggests
Eurofound’s approach borrows from the economist Carlota Perez, who treats technological revolutions as economic and social transformations rather than waves of invention. The Industrial Revolution, which began in Europe, brought rapid productivity growth, yet real wages trailed it for several decades. Labour laws, public schooling, social insurance and social dialogue came about because of industrialisation, Riso writes, not despite it.
She ends with questions rather than answers: what kind of society should AI help create, and which institutions would make that possible? Those choices, she says, cannot be left to technology companies or market forces alone, because the implications reach security, democracy, social cohesion and the future of work.
Sources
- Who benefits from AI depends on how it is governed (Eurofound blog, Sara Riso)
- Fact Sheet: President Donald J. Trump Promotes Advanced Artificial Intelligence Innovation and Security
- Why experts are worried an OpenAI model decided to hack Hugging Face
