Morning brief

Friday, July 10

0 people · 0 companies · 5 topics

TL;DR Alibaba’s Claude Code ban took effect today after the steganographic backdoor story landed on China’s NVDB. Anthropic also reset all Code usage limits following a rough week of outages and launched Reflect, a usage-analytics dashboard.

Models & research

  • model-releases Grok 4.5 first-day assessment: hallucination rate climbed from 25% to 54%, making it a cost leader at $2.49/coding task vs Claude Code’s $11.80, but unreliable for factual or high-stakes work. Artificial Analysis ranks it 4th on Intelligence Index behind Fable 5, GPT-5.5, and Opus 4.8. EU users locked out of API and app due to regulatory hold. TechTimes · Artificial Analysis

Tools & dev

  • claude-code Anthropic reset 5-hour and weekly limits for all users today following outages on July 3, 6, and 8 (including a two-hour degradation of Claude Code on the web, Cowork Remote, Claude Tag, and claude.ai on July 8). A temporary 50% weekly limit boost is due to expire July 13. Winzheng

  • Anthropic launched Reflect in Claude: a Spotify Wrapped-style usage dashboard showing topics, session frequency, and time spent over 1/3/6/12-month windows. Includes quiet hours, break nudges, and mindful-use prompts. Beta for Free/Pro/Max users with memory on. Anthropic · Dataconomy · TNW

  • agentic-ai Claude Cowork expanded to mobile and web (rolling out to Max subscribers from July 8-9). Sessions run remotely so work continues when the device is offline; scheduled tasks fire with no device active. Anthropic says more than 90% of Cowork usage is non-software work: business operations, reports, and spreadsheets. Doubled usage limits extended through Aug 5. TechCrunch · 9to5Mac

  • agentic-ai OpenAI launched ChatGPT Work on July 9: an agentic workspace mode that operates across connected apps, files, email, and desktop for hours, delivering finished deliverables from a single prompt. Available on desktop now for all plans; web and mobile rolling out to paid tiers this week. Direct answer to Claude Cowork. SiliconANGLE · Forbes · BNN Bloomberg

OSS

  • TencentCloud/TencentDB-Agent-Memory — fully local 4-tier memory pipeline for AI agents (L0 raw dialogue → L1 atomic facts → L2 scenes → L3 user persona), backed by SQLite with no external APIs. Cuts token usage by 61% and raises PersonaMem accuracy from 48% to 76% in tests with OpenClaw. Surging today with +581 stars. GitHub

Other

  • Alibaba’s ban on Claude Code and all Anthropic model products (Sonnet, Opus, Fable) took effect today; employees directed to switch to in-house Qoder. Trigger: security researchers found Claude Code versions 2.1.91–2.1.197 (April–July 1) embedded XOR-obfuscated steganographic markers in system prompt date strings to flag users from Chinese timezones and proxies. China’s NVDB issued a formal security alert on July 8. Anthropic called it an anti-abuse experiment and removed it in v2.1.198 on July 1. Tom’s Hardware · BigGo · The Register