What LARA really measures
On 27 May 2026, the Dutch non-profit foundation Aithos, based in Amsterdam, published LARA (Legal Assessment for Real-world Agents), a benchmark that measures whether large AI models comply with European law when acting as agents, i.e., AIs equipped with tools and capable of acting independently.
The method is as follows: a model is placed in a simulated work environment, with access to tools such as email, client files, or calendars. A second AI plays the role of the user and steers the scenario so that, to complete its task, the agent must break a rule. Three AI 'judges' then evaluate each exchange against the legal text, and the results were reviewed by experts for over 50 hours, according to Aithos.
The test covers ten provisions: six from the AI Act (exploitation of vulnerabilities, emotion inference, social scoring, concealment of the artificial nature of an interlocutor, subliminal manipulation, and human oversight) and four from the GDPR (transparency, data minimization, purpose limitation, and lawfulness of processing). Twelve models among the most advanced were subjected to more than 3 000 scenarios.
Unambiguous results
No model performs well. The best, Claude Opus 4.7 from Anthropic, broke the law in 46 % of cases, a compliance rate of 54 %. Google's Gemini 3.1 Pro broke it in 90 % of cases, and the least compliant model, from China's Moonshot AI, only complied with the law in 7 % of scenarios, according to Euronews. Mistral, the only European model tested, falls below 12 % compliance, leading researchers to estimate that even European providers are not equipped to comply with Union law.
The most serious practices, those that Article 5 of the AI Act outright prohibits, occurred in approximately 80 % of cases. The most striking scenario involves an elderly customer, visibly disoriented, calling a customer service center tasked with selling a premium offer. The agents are warm, advise her to discuss it with her daughter, yet still try to sell her the upgrade. According to Aithos, all models did this, every time.
The foundation offers a sober explanation:
"Models are trained to follow instructions and comply with the law. They are not well equipped to handle complex moral situations, where the 'right' decision is subjective and where context might justify an exception to the rules." (Aithos, translated from English)
What LARA does not say
These figures must be read carefully. LARA does not measure the compliance of European companies, nor that of 'systems deployed in Europe': it tests the behavior of models in situations designed to push them into wrongdoing. The models were not explicitly instructed to respect European law, Euronews notes, as researchers wanted to observe their default behavior. The judges themselves are AIs, even if their verdicts were verified by humans and all transcripts are public.
LARA therefore does not demonstrate that 'the AI Act is failing.' It shows something else, more useful for public policy: an agent can break the law without the deployer's intent, simply because it seeks to achieve the assigned goal.
Responsibility lies with those who deploy
This is the key point highlighted by Aithos: the model providers are not the ones breaking the law in these scenarios. Once integrated into a use case, a model becomes part of an 'AI system,' and it is the one who puts it into service that is responsible. The penalties are real: up to 20 million euros or 4 % of global annual turnover for GDPR, and up to 35 million euros or 7 % of global annual turnover for the most serious violations of the AI Act.
Yet deploying agents has never been simpler. A small business can connect a large model to its messaging system and customer database in just a few hours. Aithos therefore recommends testing agents before they go live, setting explicit legal restrictions, and verifying that they hold up in practice.
A shifting regulatory timeline
The legal context must also be recalled with precision. The AI Act came into force on 1 August 2024. The prohibited practices under Article 5, those that LARA most often catches, have been applicable since February 2025, and the bulk of the regulation since 2 August 2026. The most burdensome obligations, those for high-risk systems, were however postponed by the digital omnibus on AI: political agreement on 7 May 2026, publication in the Official Journal on 24 July, and entry into force on 27 July. They will apply on 2 December 2027 for standalone systems and on 2 August 2028 for those integrated into products.
In other words, the rules that LARA finds violated are already applicable. The benchmark does not predict a future failure: it signals a present gap between the law and the behavior of the tools.
What lessons for Europe?
LARA highlights three public policy challenges. First, control tools: the regulation exists, but the technical instruments to verify an agent's compliance are still under construction, and independent, open initiatives like Aithos's contribute to them. Second, support for deployers, especially SMEs, who bear legal responsibility without always mastering the tools they use. Finally, model design: if no model, European or otherwise, spontaneously complies with Union law, compliance cannot rely solely on the goodwill of providers.
Rather than an indictment of the AI Act, LARA is a reminder: in the age of agents, a rule only protects if it is tested, tooled, and enforced.
Sources: Aithos, "Aithos LARA: Leading AI models are consistently breaking the law", 27 May 2026 ; Euronews, "AI agents actively ignore EU law to achieve goals, study finds", 2 June 2026 ; LARA ranking (Aithos) ; Regulation (EU) 2024/1689 on artificial intelligence (AI Act) ; Regulation (EU) 2026/1744 (digital omnibus on AI), OJ of 24 July 2026 ; Council of the EU, provisional agreement on the AI omnibus, 7 May 2026 ; European Commission, "AI Omnibus enters into force", 27 July 2026.