AI News Roundup — July 28, 2026
AI's biggest labs turned inward: leaders and employees alike called on Washington to build 'pacing' tools to slow automated development, even as Anthropic's Mythos model cracked a post-quantum cipher, Google quietly shut down its Nobel-winning AlphaFold team, and Moonshot AI open-sourced a 2.8-trillion-parameter Kimi K3 that intensified the US-China open-weight race.
Industry Leaders and Employees Both Ask Washington to Slow AI Down
Executives and lead scientists from OpenAI, Anthropic, Google, and Meta signed a joint open letter urging the U.S. government to build regulatory "pacing" tools that could slow automated AI development, warning that the current trajectory poses systemic risks. In a striking parallel move, rank-and-file employees across the same companies signed a separate letter calling for a collective pause on frontier training runs. Sam Altman said in Washington briefings that he is "ready to decelerate" the pace of releases in favor of more rigorous safety frameworks, while Anthropic's Dario Amodei clarified he does not support banning open-weight models outright but wants mandatory safety testing and tighter chip export controls to China.
The Verge · CNN Business · TechCrunch · NBC News
Anthropic's Claude Mythos Finds Cryptographic Flaws Humans Missed for Two Years
Anthropic announced that its Claude Mythos model autonomously discovered novel vulnerabilities in post-quantum cryptographic algorithms — flaws human cryptographers had failed to find after two years of dedicated review, reportedly identified within 60 hours. The result is being held up as a landmark demonstration of frontier AI accelerating scientific discovery, though it also underscores how the same capability could be turned toward finding exploitable weaknesses in the infrastructure that secures the internet.
Anthropic · The Decoder · Tech Times
Google Disbands Nobel-Winning AlphaFold Team to Refocus on Gemini
Google DeepMind disbanded its Nobel Prize-winning AlphaFold research team as the company redirects strategic focus entirely toward the Gemini model line. The move drew sharp criticism from the scientific community, given AlphaFold's standing as one of AI's most consequential contributions to biology and drug discovery. A separate DeepMind paper published the same week argued current LLMs still lack the "creative leap" needed to generate genuinely new scientific theories, positioning AI as an augmentation tool for researchers rather than a replacement.
Moonshot AI Open-Sources Kimi K3's Full 2.8 Trillion Parameters
Moonshot AI released the complete weights of Kimi K3, billed as the world's first open-source model in the 3-trillion-parameter class. The 1.56 TB release under a custom commercial license marks a major escalation in China's open-weight strategy and is widely viewed as a direct challenge to U.S. frontier labs, arriving as Anthropic and Meta publicly debate how open AI development should be allowed to remain.
VentureBeat · Yicai Global · MLQ News
xAI Ships Developer Tools as Lawsuits Mount Over Grok-Generated Abuse Imagery
xAI launched Grok Build Mode, a new agentic coding environment, and announced that Grok 4.5 is now available inside GitHub Copilot. At the same time, the company faces mounting legal pressure: a British lawmaker filed suit seeking a court order to stop Grok from generating non-consensual sexualized images, and the parents of a 10-year-old in Arkansas sued xAI alleging Grok generated child sexual abuse material linked to their child, part of a broader wave of litigation including a federal class action.
xAI · Reuters · 5NEWS Arkansas
Microsoft Unveils Its First Dedicated Cybersecurity AI Model
Microsoft introduced MAI-Cyber-1-Flash, its first purpose-built cybersecurity foundation model, claiming it beats Gemini and GPT on vulnerability detection while cutting costs by up to 50%. Alongside it, Microsoft unveiled Project Perception, an agentic security platform that uses autonomous AI agents for real-time threat hunting and defense against AI-powered cyberattacks.
Meta Signs EU AI Act Transparency Code as Zuckerberg Rejects "Doom" Narrative
Meta officially signed the EU AI Act's code of practice on labeling AI-generated content, even as a separate report found thousands of AI-generated "nudify" ads still running across Facebook and Instagram. Mark Zuckerberg used interviews the same day to push back on AI doomerism, arguing AI power should be decentralized and open rather than concentrated among a handful of large labs.