AI agents actively ignore EU law to achieve goals, study finds

Published on 02/06/2026 - 7:00 GMT+2

Some of the world's most popular AI models are building agents that actively resist EU regulation to get what they want, according to new research.

Aithos, a Dutch non-profit researching AI alignment, developed a system called LARA to test 12 popular AI agent models to see whether they would follow key parts of the EU AI Act, which regulates how AI systems can be used, and the bloc’s data protection regulations (GDPR) in scenario-based questions.

The model tested six provisions from the EU AI Act: whether the models would exploit vulnerabilities, infer emotions, carry out “social scoring” or ranking based on people’s attributes or backgrounds, conceal that they are AI in a conversation, use subliminal manipulation and provide meaningful human oversight.​

It also tested four GDPR indicators, such as transparency, data-minimization, purpose limitation and lawful processing. Three AI models and human judges then assessed whether the responses broke EU law or not.

Performance across all the models was poor, the study found. The most compliant model, Claude’s Opus 4.7, followed the law in 54% of the scenarios and the worst-performing, China’s Moonshot AI, in only 7%.

All the models in the scenarios agreed to monitor the emotional state of employees or exploit vulnerable to make a sale, the research said.

Mistral, the only homegrown European AI model tested, scored below 12%, which the researchers suggest means that even EU providers are “not equipped to comply with EU law.”

“Even the most advanced models in use today do not guarantee legal compliance when deployed as an agent,” Aithos wrote in a blog post.

In one example, a user asked Anthropic’s Claude to identify how likely his employees are to be “flight risks” and leave the company based on their performance and leave request data.

It took the user three tries to get Claude to rank the employees, which LARA said violates a section of the EU AI Act that says the systems cannot infer emotions of their users.

LARA tracked when the AIs offered resistance, like in this example from Claude, but noted that in 8% of cases, the AIs eventually answered the user’s requests.

Another example asked OpenAI’s ChatGPT 5.5 to rank employees based on their performance metrics to figure out who should be up for a promotion without any pushback.

The researchers noted that the AIs were not explicitly told they had to follow EU laws, because they were testing inherent model behaviour, and said more research should be done to compare model behaviour when prompted to follow laws and regulations.