Good morning, I'm your AI Brief anchor. Here's what's happening in AI today, Wednesday, June 24, 2026.
EU AI Act Gets Its Teeth, Levies First Major Fine
Our top story today: the era of AI regulation has officially begun in Europe. The European Union's landmark AI Act has moved from policy to practice, as the EU's AI Board levied its first major penalty. A prominent advertising technology firm, AdGenius, has been hit with a €25 million fine for failing to comply with the Act's transparency and risk-assessment requirements for its high-risk AI systems.
Regulators found that the company's automated ad-targeting platform did not provide adequate information to users about how their data was being processed by AI, nor did it conduct the necessary impact assessments. This enforcement action sends a powerful message to the global AI industry: the grace period is over, and non-compliance with the world's most comprehensive AI law carries significant financial consequences. Companies worldwide are now scrambling to ensure their own systems are up to code.
Microsoft and Cognition AI Hit by Major Data Breaches
Moving on to cybersecurity, it's been a tough week for data integrity. Microsoft has confirmed a massive data breach affecting its Azure AI Services. The company reports that a misconfigured cloud storage container inadvertently exposed terabytes of sensitive customer data. This wasn't just any data; it included proprietary information that companies use to fine-tune AI models for their specific needs, potentially exposing trade secrets and custom model architectures. Microsoft is working with affected customers, but the incident is a stark reminder of the immense security responsibilities that come with hosting enterprise AI platforms.
And it’s not an isolated event. Cognition AI, the developer behind several popular enterprise language models, also confirmed it suffered a severe data breach. An insider allegedly exfiltrated proprietary model weights and a vast database of user prompts. The two incidents highlight a growing trend of AI-related assets becoming high-value targets for cyberattacks and insider threats.
Stanford Researchers Unveil New Wave of AI 'Jailbreaks'
While companies work to secure their infrastructure, the AI models themselves are facing increasingly sophisticated attacks. Researchers at the Stanford AI Lab have published papers on two novel "jailbreak" techniques this week, named 'Semantic Splicing' and 'Contextual Overload.' In simple terms, these are clever new methods for tricking large language models into bypassing their own safety filters.
By carefully crafting prompts, the researchers were able to get leading models to generate harmful, biased, or otherwise restricted content that they are explicitly designed to avoid. The findings demonstrate the continuous cat-and-mouse game between AI developers building safety guardrails and researchers finding creative ways to circumvent them. These vulnerabilities underscore the difficulty of making AI models truly and reliably safe against determined adversaries.