Good morning, I'm your AI Brief anchor. Here's what's happening in AI today, Friday, June 12, 2026.
Anthropic Challenges OpenAI with Record-Breaking Claude Fable 5
Our top story this morning: the AI model race is heating up again. Anthropic has just unveiled its new family of models, Claude Fable 5 and Mythos 5, and the results are turning heads. The flagship model, Mythos 5, has achieved a record-breaking score of 92.1 percent on the MMLU benchmark, a key test measuring an AI's general knowledge and problem-solving abilities.
This is the first time a commercial model has officially broken the 92 percent barrier, narrowly edging out frontrunners from both OpenAI and Google. Anthropic claims the new Fable 5 family excels at complex reasoning, creative writing, and enterprise-level tasks. While the company has been known for its focus on AI safety, this release signals a major push to compete on raw performance. The move puts direct pressure on competitors, as the industry waits to see how this new level of capability will be deployed and what OpenAI’s response will be.
White House Signs Landmark AI Regulation Bill Into Law
Moving from innovation to regulation, the White House has officially signed the 'AI Trust and Transparency Act' into law. This is a landmark moment for artificial intelligence in the United States, establishing the first comprehensive federal framework for governing high-risk AI systems.
The new law mandates rigorous testing, risk assessments, and independent audits for AI used in critical sectors like finance, healthcare, and law enforcement. Companies deploying these systems will be required to provide clear documentation on how their models work and what data they were trained on. The act also establishes a new federal office, the National AI Safety Commission, to oversee compliance and update regulations as the technology evolves. For AI companies, this means the era of voluntary commitments is over; a new chapter of mandatory accountability has officially begun.
Stanford Researchers Reveal New 'Recursive' Jailbreak Technique
But even as new laws are written, the guardrails on AI models are being tested. In a sobering new paper, researchers at Stanford's Human-Centered AI Institute have detailed a novel jailbreak technique they call the 'Recursive Embedding Attack.' This method has proven effective at bypassing the safety filters of major large language models.
In simple terms, the attack works by hiding harmful instructions within layers of complex code and data, effectively tricking the AI's safety systems into ignoring them. The researchers warn that this technique is more subtle and harder to patch than previous jailbreaks. The findings highlight the ongoing cat-and-mouse game between AI developers trying to build safe systems and researchers discovering new vulnerabilities, raising critical questions about the security of models being deployed across the globe.