✔️ China blocks Nvidia H200 exports
Beijing has halted H200 imports, despite US approval for 10 firms (including Alibaba and Tencent), to force a shift to local alternatives. A new US logistics requirement—mandating that chips pass through US territory, threatening a 25% revenue cut—has further strained supply chains. China fears potential hardware backdoors, and Nvidia’s CEO is currently in Beijing for emergency talks.
✔️ Anthropic boosts Claude Code limits by 50%
Until July 13, Pro, Team, and Enterprise users get a 50% increase in usage quotas across CLI, IDE extensions, and web/desktop versions. This stacks with last week’s 2x increase for five-hour sessions. No manual activation is needed.
✔️ Codex remote control arrives on iOS/Android
ChatGPT mobile can now monitor and control Codex sessions. Users can view terminal logs, diffs, and test results, and approve commands remotely. The connection uses a secure relay—no source code or keys leave the local environment. Requires app and macOS updates; Windows support coming soon.
✔️ arXiv to ban authors for unedited LLM content
Authors are now fully responsible for AI-generated errors, hallucinations, or “chatbot artifacts” in preprints. Violators face a one-year ban. Post-ban, they must prove peer-reviewed acceptance to regain submission rights.
✔️ GPT-5.5 and Claude Mythos pass autonomous hacking tests
The UK’s AISI tested the models on complex corporate network simulations. Claude Mythos successfully cleared advanced “The Last Ones” and “Cooling Tower” scenarios, which require multi-stage planning. Researchers note that without token limits, success rates for autonomous attacks approach 100%, exceeding current safety measurement scales.
#news #ai #ml
