Checking AI raised $250M as the EU asked who checks it

The European Commission's AI Office sent its first AI Act requests for information on October 14, and both went to frontier labs: OpenAI and Anthropic were asked about security, external evaluation and monitoring of models in use, after the AI Office judged the two companies' voluntary disclosures insufficient, with Executive Vice-President Henna Virkkunen confirming the step at the International AI Summit in Dublin (AI Weekly). OpenAI closed $122B at an $852B post-money valuation in March (OpenAI), and Anthropic raised $65B at $965B in May (Anthropic). What the letters ask for is evidence about those models produced by someone other than the lab that trained them.

Today's investment record prices that work.

The checking layer raised $250 million

Arena raised a $200 million Series B co-led by Lightspeed Venture Partners and Khosla Ventures at a $3.1 billion valuation, nearly double the $1.7 billion the AI evaluation company carried after its $150 million Series A in January (The SaaS News). Scarlet raised a $50 million Series B led by Headline, with Notion Capital new and Atomico, Creandum, Kindred Capital and Entrepreneur First returning, taking the London certification body for AI and software medical devices past $75 million raised (Tech.eu).

Two shapes of one job. Arena scores models; Scarlet certifies the devices built from them. Neither trains a frontier model, and Arena's valuation nearly doubled in nine months anyway. Both are paid for a sentence a lab cannot write about itself.

The checking layer also holds the keys

Being the place where models are tested means holding credentials for the models being tested. Fireworks AI disclosed that an attacker used internal company credentials to reach environment variables from evaluation jobs in a limited number of customer accounts, including API keys and tokens held as secrets, and said it rotated employee API keys on October 9 before revoking further user keys and rotating service account tokens on October 10 (Fireworks AI).

That is the structural cost of what the AI Office is asking for. An external evaluator is, by design, one place that holds other people's model access and other people's secrets. The letters ask about security and about external evaluation in the same request; the Fireworks bulletin is a short account of why those are one question and not two.

What the record does not say

The AI Act letters are requests for information, not findings, and nothing published today says what the AI Office will do with the answers or when. The two labs have not responded on the record. What today does show is the gap in scale between the parties: the companies being asked are carried at $852B and $965B, and the two companies raising money this month to check work like theirs priced at $3.1 billion and below.

• European Commission: first AI Act requests for information, to two labs, on security, external evaluation and monitoring (AI Weekly)
• Arena: $200M Series B at $3.1B, up from $1.7B in January (The SaaS News)
• Scarlet: $50M Series B, past $75M raised, certifying AI and software medical devices (Tech.eu)
• Fireworks AI: evaluation job environment variables reached with internal credentials, keys rotated October 9 and 10 (Fireworks AI)

If external evaluation moves from a disclosure a lab offers to a document a regulator requires, that gap is what moves first.