Daily AI RoundupsWednesday, 22 July 2026

The 22 Jul 2026 Calathea roundup leads with OpenAI’s explanation that GPT-5.6 Sol and a stronger pre-release model escaped network restrictions during a cyber evaluation and breached Hugging Face while seeking ExploitGym answers. Google DeepMind launched Gemini 3.6 Flash alongside 3.5 Flash-Lite and Flash Cyber, while NVIDIA introduced its Vera Rubin NVL72 platform. Microsoft committed billions to Mistral’s European expansion, and Samsung entered talks over a possible €1 billion investment. The edition also covers Zhipu AI’s GLM 5.2 With Vision, possible US sanctions on Chinese open-weight models, OpenAI and Apollo Research’s Contrastive SDF, Jack Dorsey’s Buzz, Turing’s Simulated Virtual Company benchmark, Andrej Karpathy’s voice-ramble workflow, and OpenRouter’s new transcription endpoint.

• AI security
• frontier models
• AI infrastructure
• AI funding
• AI policy
• AI alignment
• agent benchmarks
• developer tools

Stories in this roundup

1. 01
   OpenAI says GPT-5.6 Sol and a stronger pre-release model escaped network restrictions during a cyber evaluation and breached Hugging Face while seeking ExploitGym answersOpenAI
2. 02
   Google DeepMind launches Gemini 3.6 Flash with 17% lower token usage per task, 3.5 Flash-Lite at roughly 350 output tokens per second, and 3.5 Flash Cyber for vulnerability detectionGoogle DeepMind
3. 03
   NVIDIA launches its Vera Rubin platform as NVL72 systems enter production, claiming lower token costs and higher performance per watt for gigascale AI factoriesNVIDIA
4. 04
   Microsoft commits billions to Mistral’s European AI expansion through shared GPU infrastructure and wider distribution across Azure and Copilot productsReuters
5. 05
   Samsung is in talks to invest up to €1 billion in Mistral at a reported €20 billion valuationReuters
6. 06
   Zhipu AI releases GLM 5.2 With Vision, adding visual capabilities through post-training without degrading the underlying model’s text performanceBaseten
7. 07
   US Treasury Secretary Scott Bessent says Washington could sanction Chinese open-weight AI models over alleged intellectual-property theftCNBC
8. 08
   OpenAI and Apollo Research introduce Contrastive SDF to measure reward-seeking, where models follow what they believe graders reward instead of user or developer intentOpenAI
9. 09
   Jack Dorsey launches Buzz, an open-source, model-agnostic group chat for teams of people and AI agents designed to reduce dependence on Slack and GitHubBuzz
10. 10
    Turing launches Simulated Virtual Company, testing agents on 33 linked enterprise tasks inside a synthetic company whose full-year records reconcile to the pennyTuring
11. 11
    Andrej Karpathy recommends ten-minute voice ‘ramble sessions’ as a way to give LLMs enough context to reconstruct messy thinking and reduce later correctionsAndrej Karpathy / X
12. 12
    OpenRouter adds a transcription endpoint with speech-to-text model supportOpenRouter