Superintelligence Europe — No. 015

Pony.ai, Verne, and Uber launched commercial robotaxi service in Zagreb on Wednesday morning — the first paid, public autonomous ride-hailing service in European history. Members of the public can book and pay for rides through the Verne app, with Uber app integration to follow. The initial service zone covers approximately 90 square kilometres across central Zagreb, including the airport, with the service running daily from 7am to 9pm. The vehicles are BAIC Arcfox Alpha T5 cars equipped with Pony.ai’s seventh-generation autonomous driving system. Safety operators remain onboard during this initial commercial phase, with a transition to fully driverless operation planned subject to regulatory approval and performance metrics.

The partnership structure is a joint-deployment model: Pony.ai contributes its autonomous driving platform and operational expertise; Verne — a company spun out of Mate Rimac’s Rimac Group — owns the fleet and leads local operations and regulatory execution; Uber provides platform integration and strategic investment. The service launched just two weeks after the three companies announced their partnership, a pace that stands in sharp contrast to the multi-year delays that have characterised rival European autonomous vehicle programmes. The French Air Force Chief of Staff’s description of the Eurodrone as “yesterday’s drone we can get tomorrow” has an uncomfortable parallel in the autonomous vehicle space. Waymo is mapping London. Volkswagen’s MOIA is scaling in Hamburg. Multiple German programmes are developing. Zagreb is operating.

Amnesty International published a direct attack on the Digital Omnibus on Wednesday, framing the Commission’s simplification programme as a deregulatory push benefiting large technology companies at the expense of citizen rights. The piece — published with explicit reference to the impending trilogue — argues that proposals presented as reducing compliance burden amount to an unprecedented rollback of EU digital protections, targeting the AI Act, the GDPR, the Digital Services Act, and potentially the Digital Markets Act through the ongoing Digital Fitness Check. The intervention reflects a civil society consensus that has been building since November 2025: the organisations that spent years shaping the AI Act’s protections are watching them be reopened before they have fully come into force.

The specific AI Act provisions under threat, as identified in Amnesty’s analysis and confirmed by CDT Europe’s ongoing work, include the registration requirement for AI systems self-exempted from the high-risk category — the transparency safeguard that allows regulators to scrutinise whether providers have correctly classified their systems as non-high-risk. The Omnibus proposal would remove that requirement entirely, allowing AI companies to self-certify risk levels without any public disclosure. Amnesty’s analysis also flags proposed GDPR changes that would redefine what constitutes personal data and create new carve-outs for AI training use cases. The combined effect, the coalition argues, would make the EU’s digital rulebook materially weaker than the one that emerged from five years of legislative negotiation.

Christoph Schell, CEO of Kuka AG — the German robotics manufacturer owned by China’s Midea Group and supplying the likes of Volkswagen and Airbus — told Bloomberg on Wednesday that many of Europe’s industrial companies are moving too slowly on artificial intelligence, creating conditions for faster-moving global competitors to overtake them. The specific barrier Schell identified is not capital or ambition but infrastructure: legacy systems and a reluctance to change mean that many European factories remain disconnected and are unable to make effective use of the data they generate. AI cannot improve a factory it cannot see. And most European factories, in Schell’s assessment, are still not legible to AI at the level required for meaningful optimisation.

The intervention carries weight because of its source. Kuka is not a consultancy or a software vendor with an incentive to overstate the AI opportunity in manufacturing. It is a robotics company with 550,000 machines installed globally — more than 50 percent of global robotics demand now comes from China, where Kuka’s China revenue exceeded €1 billion for the first time in 2025. Schell’s observation about European factory AI adoption is not a prediction. It is a view from inside the production floor, shaped by Kuka’s direct commercial experience in both Asian and European markets simultaneously. The company launched its AI Automation Management Platform at NVIDIA GTC in March precisely to address the integration problem Schell is describing — connecting AI reasoning to physical robots in dynamic production environments.

Trent AI, a London-based agentic security company founded in 2025, emerged from stealth on Wednesday with a $13 million seed round led by LocalGlobe and Cambridge Innovation Capital. The round included angel participation from Joaquin Quiñonero Candela of OpenAI, Avinash Bhat of AWS, Ippokratis Pandis of Databricks, and Tony Jebara, former VP Engineering and Head of AI/ML at Spotify. The company was co-founded by Eno Thereska (CEO), Neil Lawrence (Chief Scientist, DeepMind Professor of Machine Learning at Cambridge University), and Zhenwen Dai (CTO) — all with backgrounds in AWS, Alcion, Spotify, and frontier ML research. The product is a multi-agent security platform designed to continuously scan, assess, mitigate, and evaluate risks across the code, infrastructure, and runtime environments of AI agent deployments.

The company is addressing a structural gap that has widened as enterprise AI agent adoption has accelerated faster than security frameworks have evolved. According to Deloitte’s 2026 State of AI report cited by Trent AI, nearly three in four companies plan to deploy agentic AI within two years, but only one in five report having a mature governance model for autonomous agents. Traditional security tools — designed for static applications with periodic scanning — are not architecturally suited to the continuously changing, interconnected, autonomously acting systems that constitute agentic AI. Trent AI’s four groups of agents work in orchestration: finding exploits, ranking them by severity, simulating multi-step attack paths, and validating remediation. Early design partners include Canopy, Commscentre, and Weblogic.