Anthropic sends junior engineer to face European Parliament on AI safety

A hearing intended to probe the safety implications of Anthropic's most advanced AI models descended into farce on Tuesday when the San Francisco company dispatched a recently hired engineer to face the European Parliament's internal market committee via video link from New York. Lawmakers had requested Sarah Heck, Anthropic's head of public policy. They received Donny Greenberg, a technical employee who joined the company in April through its acquisition of Runhouse and who opened his remarks by stating: "I'm a technical person, not a policy person."

A mismatch of expectations in Brussels

The session was the first opportunity for EU legislators to question Anthropic directly since the company restricted access to Mythos, a model described as capable of finding and exploiting software vulnerabilities, to a closed group of trusted American firms in early April. The European Commission, the EU's cybersecurity agency ENISA and national security bodies have been pressing for access to both Mythos and Fable, a second cyber-capable system, arguing that critical infrastructure protection depends on independent evaluation. Instead they encountered a witness who could not address the policy decisions behind that restriction.

Dutch Greens MEP Kim van Sparrentak set the tone early. "I guess you can expect a lot of policy-related questions that perhaps you won't be able to answer as a technical guy," she told Greenberg. "And I think Anthropic could have known that." Her verdict after the hearing was unsparing: "It's clear from this hearing Anthropic doesn't care about Europe." Reinier van Lanschot, another Dutch Greens MEP, said he found it "a bit disappointing that in this exchange we cannot touch on the policy side." Anna Cavazzini, the German Greens MEP who chairs the committee, warned the company to brace for a "next exchange of views we're definitely having."

Reading from the screen?

The dynamic in the room grew more strained as several MEPs noted Greenberg's eyes moving across his monitor. Spanish conservative Pablo Arias Echeverría asked directly: "Are you answering our questions, or are you just following what your AI model is telling you to say?" Greenberg replied: "I should take that as a compliment since we're very proud of Claude." The exchange encapsulated a deeper frustration: the Parliament had summoned a company making decisions with geopolitical consequences and received a staffer unprepared to defend them.

Greenberg did address technical substance. He acknowledged that cyber-capable models demonstrate AI's dual-use nature and argued that now is the moment for companies to strengthen their cyber resilience. He said Anthropic is working with the Commission's AI Office and ENISA to enable that cooperation through Glasswing, the programme designed to share Mythos with trusted cyber defenders. But on the central political question, why European institutions were excluded from the trusted circle, he had no mandate to speak.

The policy vacuum behind Mythos

Anthropic's April decision to limit Mythos to US entities landed at a sensitive moment. The EU's AI Act, which entered into force in August 2024, imposes tiered obligations on general-purpose models deemed to carry systemic risk. Mythos and Fable almost certainly fall into that category. The Commission's AI Office, established to oversee compliance, has been building its technical capacity to evaluate such systems. Denying it access to the models themselves undermines the regulatory architecture before it has fully taken shape.

The restriction also reflects a broader pattern. US frontier labs have historically treated Europe as a regulatory jurisdiction to be managed rather than a partner in safety research. OpenAI, Anthropic's chief rival, initially limited European access to its most capable models citing compliance complexity. Google DeepMind has taken a more collaborative stance, though its latest models still reach European researchers later than US counterparts. The Mythos decision, however, was explicit: a cyber-capable system withheld from the very agencies charged with protecting European critical infrastructure.

Anthropic's defence and the credibility gap

In a statement after the hearing, an Anthropic spokesperson said the company was "grateful for a frank exchange" and had fielded "one of our senior technical experts who addressed multiple substantive questions on the record." The description of Greenberg as a senior technical expert stretched credulity in the room. He joined the company three months earlier through an acqui-hire of Runhouse, a small infrastructure startup, and his public profile before April consisted of engineering blog posts. The committee had asked for the policy lead; it received a recent hire whose primary qualification was proximity to the model's code.

The company's choice suggests either a miscalculation of European political sensitivity or a deliberate decision to avoid committing to policy positions on the record. Neither interpretation flatters Anthropic's engagement strategy. The AI Act requires providers of systemic-risk models to cooperate with the AI Office, submit to independent audits and report serious incidents. A company that sends an engineer to a parliamentary hearing on those obligations signals either ignorance of the legislative timeline or indifference to it.

Contrasting approaches on the same day

The Anthropic hearing did not occur in isolation. Earlier on Tuesday, a representative of OpenAI told the same committee that the company was working to avoid staged rollouts of its new models, a phrase understood to mean simultaneous global release rather than the US-first pattern that has become standard. If delivered, that would represent a shift from OpenAI's approach with GPT-4, which reached European users months after its American debut. The company offered no timeline or mechanism for the commitment.

Meanwhile, Google DeepMind chief executive Demis Hassabis used a separate Brussels engagement to call for the United States to create an industry-funded standards body to test frontier AI models and set benchmarks. Hassabis argued that voluntary industry cooperation, backed by government authority, could move faster than treaty-based regulation. The proposal aligns with the UK's AI Safety Institute model but places the institutional home in Washington, not London or Brussels. It also presupposes US legislative action that shows no sign of materialising before the 2026 midterms.

What the episode reveals about EU leverage

The hearing exposed a structural asymmetry. The EU possesses the world's most comprehensive AI regulatory framework and a market of 450 million users. Yet the companies building the most powerful systems remain overwhelmingly American, answerable to US capital and US national security priorities. When Anthropic decides that Mythos stays in American hands, the EU's legal tools, the AI Act's systemic-risk provisions, the Cyber Resilience Act, the proposed Cyber Solidarity Act, cannot compel access retroactively. They can only penalise non-compliance after the fact.

European officials understand this. The AI Office has been negotiating voluntary codes of practice with frontier labs since late 2024. The Mythos restriction suggests those negotiations have not produced binding commitments on model access. Cavazzini's promise of a further exchange of views is not a threat; it is an acknowledgement that the Parliament's formal powers, hearings, reports, budgetary oversight of the AI Office, are the only levers available short of enforcement proceedings that would take years.