The afternoon briefing.

AI Security Risks. White House export restrictions on Anthropic's Mythos were reportedly influenced by fears of Chinese access, raising serious national security concerns. This potential breach highlights the vulnerability of advanced AI models to state-sponsored actors and the risks of reverse engineering through distillation. Meanwhile, KPMG pulled an AI adoption report after it was found to contain fabricated case studies, underscoring the growing problem of "secondary hallucinations" and misinformation in the AI industry.

AI Agent Deployment. The rise of autonomous AI agents is fundamentally reshaping enterprise security, with many organizations struggling to keep pace with new attack vectors. Google Cloud introduced its Open Knowledge Format, standardizing organizational knowledge into Markdown files for better AI agent consumption. This move aims to make internal data more portable and usable for AI systems, addressing a critical need for structured information in agent-driven workflows.

Frontier Model Performance. Xiaomi's MiMo Code claims to outperform Claude Code in extended coding tasks, specifically past 200 steps, indicating progress in agent endurance for complex software development. However, a "homegrown" LLM from Rio de Janeiro's city government, Rio3.5, which recently beat Qwen3.7 in benchmarks, appears to be a merge of an existing model rather than a novel creation. This suggests a trend of leveraging and adapting existing foundational models for specific applications.

Workforce and Society. A recent report indicates that nearly half of workers fear AI will take their jobs, even as the use of AI agents in companies has soared by 90% in the past year. This growing anxiety highlights the need for reskilling and adaptation in the workforce, as AI continues to integrate into daily operations. Experts suggest that adaptability and critical thinking will be the most in-demand qualities for job-seekers in the evolving AI era.