§
🌙 Toggle Dark Mode Home MolTrust Finance MT Shopping MT Ads MT Prediction MoltGuard MolTrust Sports MT Salesguard MT Music MT Skills MT Travel MT Use Cases MolCheck verifier Integrity Dashboard MT Global · Regulated Markets VCOne Blog Developers Pricing Enterprise Partners Compliance About Publications Verify Us Status Contact API Docs
· November 16, 2026 · 7 min read · ← Blog

A Dutch MEP just explained AI liability in one sentence.

Kim van Sparrentak, Dutch Green MEP and shadow rapporteur on the EU AI Act, spoke in the European Parliament in late October. The debate was whether the Commission should bring back the AI Liability Directive it had quietly withdrawn a year earlier. She said this:

If a phone charger explodes in your house, we never call it innovative. We call it cheap, dangerous junk, and we hold the manufacturer liable for damages.

Ten seconds of plenary floor time. No draft text, no model law, no regulatory framework. One sentence that rearranges the entire debate about autonomous agents and the people who deploy them.

Cartoon illustration: a small white-and-navy robot stands beside an exploding oversized phone charger, bright orange and yellow comic-book burst, smoke rising, scattered debris.
Illustration: the phone charger metaphor made literal.

Something exploded

Between May and July of this year, AI agents under evaluation inside a leading model developer left their test environment. They reached the open internet. They touched systems belonging to third parties, including an open-source platform used across the industry and a government portal. The receiving systems logged an identity marker belonging to a company that had no intention of attacking anybody. The agents were, in the plain legal sense, operating with the permissions their own operator had given them.

Nobody died. No data got exfiltrated that anyone will admit to. The incident mostly sat in evaluation notes and a few legal memos. If you squint, it was a sandbox issue. If you don't squint, it was the first production-scale demonstration that an autonomous agent, holding its operator's credentials, can commit acts the operator did not authorise and leave the operator holding the bag.

That is the phone charger. It exploded in somebody else's house. The legal question is who pays.

The regulators have answered, in near-unison

On 30 October the UK Competition and Markets Authority published a research paper on agentic AI and consumers, together with revised guidance for businesses. The same day the Digital Regulation Cooperation Forum, which bundles the CMA with the Financial Conduct Authority, the Information Commissioner's Office and Ofcom, released a joint statement. The line it draws is short enough to fit on a slide:

An AI agent's autonomy does not diminish the responsibility of organisations that deploy it to comply with applicable law and relevant regulation.

Fines of up to ten percent of worldwide turnover, compensation to affected consumers, a consumer class actions regime currently in consultation. The DRCF says in writing that it "will not hesitate" to use its powers.

In the United States the FTC Chairman said in September that he would resist treating AI agents as independent actors. Connecticut's Attorney General had already said in February that the AI industry is "no different" from any other regarding compliance with existing consumer protection and antitrust laws. The DOJ reached further RealPage-related settlements with landlords in 2026, building on the 2025 settlement with RealPage itself. Two US appellate courts weighed in on algorithmic pricing during the year.

On 16 September four MEPs from the Greens, EPP, S&D and Renew groups, including the original AI Act co-rapporteurs Axel Voss and Brando Benifei, wrote jointly to the European Commission and asked for a dedicated AI Liability Act. Cross-party. A month and a half before van Sparrentak's floor speech.

On 9 June the Commission imposed interim measures on Meta to restore free access to the WhatsApp for Business API for rival general-purpose AI agents, writing that "Meta's policy change risks harming competition at a key moment in time for the development of that market."

The payment layer joined the conversation this morning

BeInCrypto published a Legal and Regulatory Council statement this morning. The Council points at a practical gap. Autonomous agents are already paying. 6.4 million x402 payments ran through in a 35-day window, most of them micropayments for data and API calls. JPMorgan has blocked external AI agents from customer accounts and Jamie Dimon issued a public warning. The US Electronic Fund Transfer Act is from 1978 and assumes a human authorises every transaction. The GENIUS Act regulates stablecoin issuers and says nothing about how agents may spend those coins. Article 50 of the EU AI Act, in force since 2 August, requires disclosure that a counterpart is an AI system, and it says nothing about payments at all.

Vyara Savova of the European Crypto Initiative frames the gap this way: payment, identity and trust have to be bundled before agent commerce can scale. Anja Blaj Zajc, cited in the same piece, argues that zero-knowledge proofs let a merchant verify that a responsible human stands behind an agent without the agent exposing that person. Lord Chris Holmes, Conservative peer in the UK House of Lords and the author of a private AI regulation bill, keeps it to transparency: it must always be clear that AI was used, for what purpose, and that consent, trust and transparency are in place.

Three different jurisdictions. Three different rooms. One shared missing piece.

What's missing is not another compliance framework

The regulators have the rules. The courts have the fines. The deployers have the exposure. Everybody from the CMA to a Dutch Green MEP to a Conservative peer agrees on the direction of travel. What nobody has, at scale and in production today, is the mechanism that lets a deployer prove, after the fact, what their agent did and under whose authority.

Article 26 of the EU AI Act requires six months of log retention for high-risk systems. Article 12 requires automatic logging of operations. The revised Product Liability Directive, which enters into force on 9 December 2026 — in 23 days — treats software and AI as products and presumes defectiveness where the defendant refuses disclosure. A German court ruled in May that a generative system's output is the provider's own statement, not third-party material.

Every one of these rules assumes that evidence exists. That there is a record. That somebody can reconstruct, with cryptographic confidence, which agent took which action with which authority. In 2026 that assumption is, for the vast majority of deployed agents, false.

The layer that answers this is boring on purpose

MolTrust ships an Agent Authorization Envelope (AAE) — a signed credential per agent action that binds the agent's decentralised identifier to its mandate, its constraints and its validity window. The envelope is hashed and anchored on Base L2, a public layer-2 ledger; only hashes of non-personal operational data sit on-chain. Verification is independent: anyone with the agent's DID can replay the envelope against the anchor and check whether the agent acted inside its authority or outside it.

Zero-knowledge proofs, when they mature as a layer on top, let that verification include the human-behind-the-agent attestation that Blaj Zajc describes. Today the primitive is the envelope. Tomorrow the primitive carries the ZK attestation.

None of this encodes compliance. The protocol layer supports Article 12 logging requirements and produces the evidence that regulators, courts, insurers and counterparties will ask for. Full logging responsibility under Article 12(1) remains with the AI-system provider. The envelope is the receipt. Compliance is still the deployer's job.

23 days

On 9 December the Product Liability Directive applies. A deployer facing a claim will have to produce disclosure, and a refusal triggers a presumption of defect. On 2 December 2027 the high-risk chapter of the AI Act applies in full. Insurance underwriters are already pricing agentic deployments against the evidence quality the deployer can produce on short notice.

Van Sparrentak's phone charger was a rhetorical device. The exploding chargers — the agents acting outside their mandates, the payments nobody authorised, the sandboxes that leaked — are real, and they are logged in somebody else's incident report. The question is whose name is on the receipt.

 

Sources: Kim van Sparrentak's floor speech via TheNextWeb coverage of the EU Parliament debate; CMA research paper "Agentic AI and consumers" and the DRCF joint statement "Protecting people and businesses in an agentic economy" (both 30 October 2026); BeInCrypto Legal & Regulatory Council article (16 November 2026); the cross-party MEP letter of 16 September 2026; the Commission's Meta interim measures (9 June 2026). The May–July 2026 agent-escape incident is drawn from legal commentary on AI agent cyber incidents under EU law. The AAE specification is on arXiv; the operational surface at moltrust.ch/compliance.