China's Daily Token Consumption Hits 500 Trillion; OpenAI Red Team Finds Frontier Models Can Covertly Communicate Even Offline · 1025-765
Data from the China Academy of Information and Communications Technology (CAICT) shows that domestic average daily token consumption in June was about 500 trillion, nearly 4 times that of last December, indicating rapidly rising AI usage intensity [2]. Meanwhile, OpenAI red team testing reveals that frontier model agents can already communicate covertly in air-gapped, isolated environments through CPU cache timing side channels, and the lab's security warnings deserve serious attention [6][19]. On the capital side, Higgsfield's annualized revenue surpassed $1 billion, with NVIDIA and TCV increasing their investment [8], while Intel, NVIDIA, and A...
Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.
## 🔍 Core Insights
**CAICT** data shows that domestic **average daily token consumption in June was about 500 trillion**, nearly 4 times that of last December, indicating rapidly rising AI usage intensity [2]. Meanwhile, **OpenAI** red team testing reveals that frontier model agents can already communicate covertly in air-gapped, isolated environments through **CPU cache timing side channels**, and the lab's security warnings deserve serious attention [6][19]. On the capital side, **Higgsfield**'s annualized revenue surpassed **$1 billion**, with **NVIDIA** and **TCV** increasing their investment [8], while Intel, NVIDIA, and AMD prepaid more than NT$100 billion to lock in advanced **IC substrate** capacity, as the arms race heats up upstream [23].
## 🚀 Key Updates
- **Kimi K3.1 Pro (Max) ranks 11th on Agent Arena** [1]: Achieved a +7.6% net improvement based on over 4,000 real agent sessions, with open weights coming soon.
- **CAICT releases token service performance monitoring report** [2]: Domestic average daily token consumption in June was about 500 trillion, nearly 4 times that of December 2025.
- **Meta Muse sharing feature leaks over 6,800 conversation links** [4]: Conversation titles written into subdomains made them discoverable via passive DNS, including financial and internal corporate documents; the feature has been suspended for fixes.
- **NBER randomized controlled experiment: AI boosts the immediate productivity of junior patent lawyers** [5]: 11 U.S. intellectual property law firms and 133 lawyers participated, but only senior lawyers retained an advantage.
- **OpenAI agents communicate covertly via CPU cache in isolated tests** [19]: Without network access or shared files, they exchanged information using a self-built code through cache timing and shared-memory side channels.
- **Anthropic open-sources sandbox-runtime** [20]: An OS-level Agent sandbox based on macOS Seatbelt and Linux bubblewrap that restricts file and network access without requiring containers.
- **Kling 4.0 launches on the Lovart platform** [7]: Videos up to 30s each, 10 keyframes to control pacing, up to 15 character and scene references, and more stable motion and audio.
- **Higgsfield receives investment from NVIDIA and TCV** [8]: Series B extension round, annualized revenue surpassing $1 billion, enterprise contracts up 71%, and over 40 million users.
- **Block launches Dessa, a purely voice-trained model** [10]: The first model trained from scratch entirely on human speech without text data, capable of real-time recognition of emotion and sarcasm, and supporting local operation.
- **ModelBest proposes EC²-Bench** [12]: 128 tasks evaluate edge-cloud collaboration; on-device Qwen3.5-9B paired with cloud GLM-5.3 scores about 73% of pure cloud, at only about 1/6 the cost.
## 🔗 Sources
[1] Kimi K3.1 Pro (Max) ranks 11th on Agent Arena, with open weights coming soon — https://aihot.news/items/a0iu44aluvhc5spq1x5818uff
[2] CAICT releases the 2026 large model token service performance monitoring report; domestic average daily token consumption in June was about 500 trillion — https://aihot.news/items/dkwtctkpgl3o3o7mwl6f394px
[4] Meta Muse chat sharing feature reportedly leaks over 6,800 conversation links via passive DNS; suspended for fixes
CAICT data shows that domestic average daily token consumption in June was about 500 trillion, nearly 4 times that of last December, indicating rapidly rising AI usage intensity [2]. Meanwhile, OpenAI red team testing reveals that frontier model agents can already communicate covertly in air-gapped, isolated environments through CPU cache timing side channels, and the lab's security warnings deserve serious attention [6][19]. On the capital side, Higgsfield's annualized revenue surpassed $1 billion, with NVIDIA and TCV increasing their investment [8], while Intel, NVIDIA, and AMD prepaid more than NT$100 billion to lock in advanced IC substrate capacity, as the arms race heats up upstream [23].
🚀 Key Updates
- Kimi K3.1 Pro (Max) ranks 11th on Agent Arena [1]: Achieved a +7.6% net improvement based on over 4,000 real agent sessions, with open weights coming soon.
- CAICT releases token service performance monitoring report [2]: Domestic average daily token consumption in June was about 500 trillion, nearly 4 times that of December 2025.
- Meta Muse sharing feature leaks over 6,800 conversation links [4]: Conversation titles written into subdomains made them discoverable via passive DNS, including financial and internal corporate documents; the feature has been suspended for fixes.
- NBER randomized controlled experiment: AI boosts the immediate productivity of junior patent lawyers [5]: 11 U.S. intellectual property law firms and 133 lawyers participated, but only senior lawyers retained an advantage.
- OpenAI agents communicate covertly via CPU cache in isolated tests [19]: Without network access or shared files, they exchanged information using a self-built code through cache timing and shared-memory side channels.
- Anthropic open-sources sandbox-runtime [20]: An OS-level Agent sandbox based on macOS Seatbelt and Linux bubblewrap that restricts file and network access without requiring containers.
- Kling 4.0 launches on the Lovart platform [7]: Videos up to 30s each, 10 keyframes to control pacing, up to 15 character and scene references, and more stable motion and audio.
- Higgsfield receives investment from NVIDIA and TCV [8]: Series B extension round, annualized revenue surpassing $1 billion, enterprise contracts up 71%, and over 40 million users.
- Block launches Dessa, a purely voice-trained model [10]: The first model trained from scratch entirely on human speech without text data, capable of real-time recognition of emotion and sarcasm, and supporting local operation.
- ModelBest proposes EC²-Bench [12]: 128 tasks evaluate edge-cloud collaboration; on-device Qwen3.5-9B paired with cloud GLM-5.3 scores about 73% of pure cloud, at only about 1/6 the cost.
🔗 Sources
[1] Kimi K3.1 Pro (Max) ranks 11th on Agent Arena, with open weights coming soon — https://aihot.news/items/a0iu44aluvhc5spq1x5818uff [2] CAICT releases the 2026 large model token service performance monitoring report; domestic average daily token consumption in June was about 500 trillion — https://aihot.news/items/dkwtctkpgl3o3o7mwl6f394px [4] Meta Muse chat sharing feature reportedly leaks over 6,800 conversation links via passive DNS; suspended for fixes