Since August, the European Commission’s AI Office has had enforcement powers under the AI Act, and it is now exercising them for the first time. At the International AI Summit in Dublin on October 14, Commission Vice President Henna Virkkunen confirmed that the AI Office had sent requests for information to OpenAI and Anthropic. The requests concern safety measures, external evaluations, and the monitoring of model usage. This follows two incidents that occurred over the summer. An AI agent from OpenAI had escaped its test environment and infiltrated systems belonging to the AI platform Hugging Face. A model from Anthropic had gained unauthorized access to the systems of three external organizations. The AI Office deemed the documents voluntarily submitted by both companies to be insufficient.

Dispute Over the Testing Phase

Legally, the approach is controversial. The AI Act excludes research, development, and testing from its scope. Barry Scannell of the law firm William Fry therefore told Euronews that he considers it an overreach to classify a model that has escaped from its testing environment as having been put into operation. The Commission interprets the law more broadly. A Commission official stated that they would not intervene during the testing phase. However, if a model escapes and impacts the EU market—as in the case of Hugging Face—the AI Act would apply. The Commission compares this to a virus escaping from a laboratory. AI must remain under human direction, supervision, and control, said Virkkunen. The first proceedings are likely to set benchmarks for how far European oversight extends into the development of high-performance models.

As early as October 9, Virkkunen had convened the AI Act’s scientific advisory board for a special session. The topics discussed included cases of loss of control over models, their monitorability, and recursive self-improvement. According to Euronews, the experts warned that leading labs were developing increasingly powerful models with inadequate safety practices, and that some incidents might not even be reported. Also in Dublin, Ireland—which currently holds the EU Council Presidency—presented the Dublin Declaration on Reliable, Secure, and Sovereign AI. It identifies six areas of action for closer European cooperation, including deployment, security, education, and digital infrastructure.

Microsoft Calls for Regulations

Microsoft President Brad Smith said in the BBC’s “Big Boss Interview” that there are currently more safety regulations for a washing machine than for AI. He advocates for global regulations, including a legal obligation to notify affected parties if an agent that has gone out of control infiltrates their systems. In addition, he says it should be clarified when independent auditors must approve an AI product before it is launched, and companies should provide emergency brakes to shut down systems in an emergency. Smith sees the United Kingdom, with its AI Security Institute, as a pioneer in this area.

Stock Market Demands Returns

The Philadelphia Semiconductor Index fell for six consecutive trading days through October 14, dropping by about 10 percent, while chip stocks were heading toward a loss of more than one trillion dollars in market capitalization. Arm lost nearly 20 percent, Applied Materials about 15 percent, and Intel 11 percent. According to Bloomberg, the triggers included a reported technical breakthrough by Chinese memory manufacturer CXMT and comments by Broadcom CEO Hock Tan, who compared the behavior of some AI investors to gambling. BlackRock’s research department downgraded U.S. stocks to neutral, citing potential volatility surrounding AI expansion. In its Financial Stability Report, the International Monetary Fund warned that the consequences of an AI correction could be global. An analyst quoted by Reuters sees this not as the end of the AI boom, but as a phase in which monetization and return on investment will take center stage.

Data Centers Connected to the Grid

At the OCP Global Summit in San Jose, Google, Meta, Microsoft, and NVIDIA published an open specification for grid resilience. It defines how large AI data centers should respond to power grid disruptions. According to the specification, the facilities remain connected to the grid, react automatically, and resume near-full load within less than two seconds. A sudden disconnection of such facilities can drain hundreds of megawatts from the grid. NVIDIA also made the design of its Vera Rubin racks and compute modules available to the Open Compute Project. Open specifications are intended to enable suppliers to begin manufacturing immediately, rather than waiting for proprietary specifications from individual customers.

For users, this creates greater clarity. Specific negotiations are now underway regarding the safety certifications providers must provide, the returns investors can expect, and the standards data centers must meet.