1.
Anthropic released Claude Fable 5, the first model in its Mythos class, which led multiple benchmarks including SWE-bench Verified at 95 percent, cost about twice as much as Opus 4.8, and applied strict safety filters that blocked roughly nine percent of requests while instituting a 30-day data retention policy even for zero-data-retention contracts.
2.
Anthropic's security team found that its Mythos Preview model converted security patches for Firefox and the Windows kernel into working exploits within hours for a few thousand dollars and without specialized knowledge, completing eight attack chains before Microsoft's auto-updates had reached any device.
3.
Google upgraded NotebookLM to run on Gemini 3.5 Flash, provided it with a cloud computer for code execution, and added agent-based research with autonomous source-finding via Google Search; internal tests showed the new system outperformed the previous version up to 78.2 percent of the time.
4.
Google released DiffusionGemma, a 26-billion-parameter open model that generated text via diffusion from noise instead of autoregressive token-by-token methods; Nvidia reported it ran at roughly 1,000 tokens per second on a single H100 GPU, though output quality was lower and Google positioned it as an experimental tool for developers.
5.
Germany's National Security Council approved creating a DE-AISI AI Safety Institute modeled after the UK's AISI to test frontier models from providers such as Anthropic and OpenAI for security risks, and stated the EU remained dependent on U.S. and Chinese AI technology in the absence of its own frontier models.
