OpenAI gives EU GPT-5.5-Cyber access

Cybersecurity models are becoming a bargaining chip in Europe — not just a product. The immediate problem is simple: regulators and defenders want to test the strongest systems before attackers do. But the gap has been awkward. The EU can write rules for frontier AI, yet it still needs the companies building those models to voluntarily open the door. (cnbc.com) This week, OpenAI did that in Europe by offering GPT-5.5-Cyber access to EU institutions and a wider set of European companies. (channelnewsasia.com) ### What is GPT-5.5-Cyber? It is a security-focused variant of OpenAI’s GPT-5.5, built to be more permissive on cybersecurity tasks for vetted users. OpenAI started a limited preview on May 7 for approved security teams, framing it as a model for defensive work like finding vulnerabilities and helping patch systems rather than a general public release. (the-decoder.com) ### What changed in Europe? The new part is direct EU access. OpenAI told CNBC on May 11 that the European Union would get access to the model, and an EU Commission spokesperson confirmed talks were underway this week over how that access would work in practice. The offer is not just for Brussels staff — OpenAI also said European businesses, governments, cyber authorities and the EU AI Office would be included in the access plan. (cnbc.com) ### Which companies are getting it? OpenAI is also widening distribution through its Trusted Access for Cyber program. On May 12, it said Deutsche Telekom, BBVA and Telefónica were among the European companies getting access, alongside dozens of others in financial services, telecoms, energy and public services. Basically, these are sectors where a software flaw can turn into a national problem fast. (cnbc.com) ### Why does Brussels care so much? Because frontier cyber models are weirdly dual-use. The same system that helps defenders find and fix bugs can also help somebody chain those bugs into an attack. So the EU’s problem is not only “how do we regulate this?” It is also “how do we inspect it early enough to matter?” OpenAI’s offer gives Brussels a practical answer, at least for one major lab. (cnbc.com) ### Why is Anthropic part of the story? Because the comparison makes the politics obvious. Anthropic’s Mythos set off a lot of the current alarm around AI-enabled cyber capability, but the Commission still had no equivalent access after four to five meetings with the company as of May 11. That means Europe’s regulators are dealing with two realities at once — formal authority on paper, and dependence on voluntary cooperation in practice. (channelnewsasia.com) ### Is this mainly about safety or market access? Both. Access now works like trust, but also like leverage. If a company gives European institutions visibility into a powerful model, that can ease relationships with regulators and major customers at the same time. And if it withholds access, it risks looking harder to supervise just as Europe is deciding which labs it can work with closely. (the-decoder.com) That is an inference from the timing and the structure of these deals, but it fits the pattern in Brussels this week. ### What is the catch? The catch is that none of this removes the underlying dependency. Europe still does not have an automatic way to examine every frontier model that matters. It has one company volunteering access, another still hesitating, and a policy system racing to catch up with tools that are already useful enough to change cyber defense operations. (the-decoder.com) ### Bottom line? OpenAI did more than hand Europe a model. It showed that in frontier AI, controlled access is becoming part of regulation, diplomacy and enterprise sales all at once. And right now, that makes cooperation itself a competitive advantage. (the-decoder.com) (cnbc.com)