OpenAI forgoes full release of new models with dangerous capabilities

The decision follows a debate in the industry about the responsible handling of increasingly capable AI systems. Competitors such as Anthropic and the platform Hugging Face are also discussing similar thresholds. Observers see this as a signal that providers are weighing the potential misuse of their models more heavily than they did just a few years ago.

Security researchers regard the move as understandable, but point out that mere voluntary commitments are not sufficient. They call for transparent testing, independent audits, and clear reporting channels to authorities before new capability levels are reached. Only in this way, they say, can it be verified whether a provider is actually applying its own criteria.

Security researchers call for independent review

In practice, the policy means that OpenAI could in future roll out certain internal models only in a restricted variant or behind additional safeguards. These include, for example, limited API access, stronger usage filters, and mandatory security checks for developers who want to access the models.

At the same time, the company stresses that research into more capable agents will continue. It wants to make the benefits of the technology usable without letting tools for large-scale attacks fall into the wrong hands. A spokesperson referred to the principle of releasing new capabilities gradually.

This stance is in line with demands from governments in the US and in Europe, which want to make AI safety a fixed component of product development. The EU is anchoring corresponding obligations in the AI Act; in the US, consultations between supervisory authorities and manufacturers on voluntary standards are underway.

Regulatory framework in the EU and US

The line is also causing discussion within OpenAI itself. Parts of the workforce argue that restrictions imposed too early would hamper innovation, while safety teams push for clear red lines. According to management, balancing the pace of research against risk limitation remains an ongoing task.

For the coming months, the company announced that it would regularly update its safety guidelines, incorporating external experts as well. Industry observers such as the news agency Reuter view the stance as an indicator that safety considerations have arrived in commercial AI development.

Regardless of the self-restriction, it remains unclear how robust the "critical capabilities" threshold actually is. The classification depends on internal tests whose methodology the company has so far only partially disclosed. External experts caution that, without independent verification, there is a risk that models could be released just below the threshold without the consequences being fully assessable.

Open questions about the threshold's robustness

One thing is certain: with its current stance, OpenAI is taking on a pioneering role among the major AI providers. Whether other manufacturers such as Anthropic or open-source platforms such as Hugging Face follow suit is likely to significantly shape the debate on safe AI development over the coming quarters.

Industry insiders also see the step as an indirect indication of the ongoing race to develop ever more capable AI agents—that is, AI systems that independently carry out tasks on a computer. As soon as such agents master complex attack patterns, the pressure will mount on all providers to adapt their protective measures.