European Union regulators have made little progress in negotiations with Anthropic over testing the company's AI system, Mythos. The standoff reveals a cybersecurity gap that Brussels has struggled to close — and could push the bloc toward stricter rules and more homegrown AI development.
Why the talks broke down
EU officials asked to run safety evaluations on Mythos before it is deployed widely in Europe. Anthropic pushed back, citing trade secrets and competitive concerns. Several rounds of discussion failed to produce a workable framework for access. Neither side has described the specific sticking points, but people familiar with the matter say the core issue is how much of the system's inner workings regulators get to see.
This isn't a new problem. Brussels has faced similar resistance from other AI developers. But Mythos is a high-profile case — it's a large model Anthropic has marketed for use in finance, health care, and government services. Without testing, EU watchdogs can't verify claims about safety or bias.
The cybersecurity dimension
The difficulty in getting to Mythos points to a broader vulnerability. If the EU can't inspect a foreign company's AI, it can't know if the system has backdoors or could be exploited by hostile actors. The situation highlights a gap in the bloc's digital defenses. Officials privately acknowledge they lack the technical infrastructure to audit advanced AI models thoroughly. That leaves the EU reliant on companies' own disclosures — a system critics have called inadequate.
Anthropic has said it conducts its own testing and shares results with governments. But EU regulators want independent verification. The standoff underscores how quickly AI is outpacing traditional regulatory tools.
What AI sovereignty means for Brussels
The deadlock is fueling a push for what officials now call “AI sovereignty.” The idea is straightforward: Europe should build its own high-capability AI systems and the testing facilities to go with them. That way the bloc wouldn't have to bargain for access to foreign models. Several EU member states have already started funding domestic AI research. The European Commission is looking at creating a joint testing lab that could certify AI models before they reach the market.
Changes to the EU's AI Act are also being discussed. The law already mandates risk-based oversight, but the Mythos case shows how easily that can be sidestepped when a developer refuses to cooperate. Some lawmakers want to add provisions that would require companies to grant real-time access to model weights and training data under certain conditions. That would face legal challenges — and fierce lobbying from industry.
The unresolved question is what happens next. Brussels has limited options. It could push for a binding resolution through the European Parliament, but that would take months. It could also impose restrictions on Mythos's use in the EU, effectively banning it until access is granted. Anthropic's next move will determine whether this becomes a test case for AI regulation or a quiet concession. Either way, the clock is ticking.