Good morning, I'm your AI Brief anchor. Here's what's happening in AI today, Monday, April 13, 2026.
Cognition Corp Suffers Catastrophic Data Breach
Our top story this morning: Emerging AI leader Cognition Corp is in crisis mode after announcing a catastrophic security breach over the weekend. The company confirmed that attackers have stolen the proprietary model weights for its flagship "Insight-5" large language model. Think of model weights as the AI's secret sauce—the core intellectual property that makes the system work.
This is a devastating blow for the highly-valued startup. In addition to the model itself, attackers also exfiltrated terabytes of user prompt data, raising serious privacy concerns for its enterprise clients. The breach effectively puts Cognition’s core technology out in the open, potentially wiping out its competitive advantage overnight.
Security analysts are calling this one of the most significant intellectual property thefts in the history of the AI industry. The incident sends a chilling message to every AI lab in the world: if your most valuable asset is digital, it can be stolen. Cognition Corp's stock is expected to plummet when markets open, and the fallout will likely be felt across the entire sector as investors re-evaluate security risks.
US and EU Announce Landmark AI Safety Pact
Moving on, and in a direct response to growing security and safety fears, governments are taking action. In a landmark announcement from Brussels, the United States and the European Union have jointly established the Transatlantic AI Safety and Certification Pact, or TAS-CERT.
This binding agreement is the first of its kind, creating a harmonized regulatory framework for AI safety across both continents. Under the new pact, AI companies wanting to operate in the US and EU will need to submit their high-risk models to a unified certification process. This process will involve rigorous third-party audits for safety, security vulnerabilities, and ethical alignment.
The goal of TAS-CERT is to prevent the exact kind of high-stakes failures we're seeing today. By setting a common standard, officials hope to streamline compliance for developers while ensuring a high bar for public safety. For businesses, this means one set of rules to follow, but it also means significantly more oversight before powerful AI systems can be deployed. The first certification requirements are set to roll out in early 2027.
Berkeley "Copy-Paste" Bot Exposes Flaws in AI Testing
And finally, a story that’s making us question how we measure AI progress at all. Researchers at UC Berkeley have developed a shockingly simple "copy-paste" bot that scored a 97 on a major AI benchmark, soundly beating sophisticated models like GPT-4o.
So, how did it work? The bot didn't reason or understand the questions. Instead, it simply searched its database for the most similar question-and-answer pair it had seen before and copied the answer. The stunning result suggests that many of our industry-standard tests may not be measuring true intelligence, but rather a model's ability to memorize and regurgitate its training data.