Daily Brief Archive

Anthropic plans October IPO targeting over $60 billion in funding and $190 billion in annualized revenue; Claude discovers 20-year Linux kernel vulnerability in 90 minutes, showcasing exponential growth in AI security offensive and defensive capabilities; Chroma releases 20B agentic search model Context-1 with inference cost just 1/25 of GPT-5

Anthropic opened a research preview of Claude computer control to Pro and Max users, limited to macOS; Mistral released the 4B-parameter Voxtral TTS with support for 9 languages and API pricing of $0.016 per thousand characters; Shield AI completed a $2 billion financing at a $12.7 billion valuation and is moving ahead with the acquisition of Aechelon

SAP plans to acquire Reltio in Q2/Q3 2026 and integrate it into Business Data Cloud to strengthen AI-ready master data; Anthropic confirms testing stronger model Mythos following unprotected storage leak and opens early access to select customers; OpenAI upgrades Responses API with built-in agent execution loop and adds Shell and containerized workspace

Google DeepMind releases Gemini 3.1 Flash Live scoring 90.8 on ComplexFuncBench Audio with SynthID watermarking on generated audio; European Parliament advances delay of AI Act high-risk compliance to December 2027 and proposes ban on 'undressing/nudification' deepfake systems; Shield AI completes $1.5B Series G at $12.7B valuation and plans to acquire Aechelon to strengthen high-fidelity simulation capabilities

Arm launches 3nm AGI CPU with projected $15 billion annual revenue in about five years; SK hynix files confidential ADR application with SEC and advances HBM expansion along with ~$7.97 billion in ASML equipment procurement; Google TurboQuant claims up to 6x KV cache compression and up to 8x attention speedup on H100 without requiring retraining

OpenAI announces shutdown of Sora video generation app and API; Alibaba DAMO Academy releases XuanTie C950, achieving SPECint2006 single-core score above 70 and claiming native support for billion-parameter large models; LiteLLM 1.82.8 suffers supply chain poisoning, triggering credential theft upon installation, requiring immediate key rotation

OpenAI brings on a former Meta advertising executive to lead the ChatGPT advertising team, testing a CPM of about $60 with a minimum ad spend commitment of $200,000–$250,000; Google Threat Intelligence uses Gemini agents to analyze over 10 million dark web posts daily, claiming 98% accuracy; Luma AI launches the autoregressive image model Uni-1, scoring 0.51 on RISEBench and pricing 2K image generation API at about 9 cents

Musk unveils Terafab chip plan, targeting 1TW annual capacity and deploying primary compute in orbital data centers; NVIDIA begins GPU deliveries to AWS on March 22, aiming to deploy over 1 million units by 2027 and co-build Ceiba supercomputer (20,736 B200 GPUs); WeChat launches ClawBot plugin, allowing OpenClaw agent integration in WeChat 8.0.70 via chat interface and QR code pairing

Cursor apologizes for failing to disclose Kimi K2.5 as the base model for Composer 2 and confirms licensing chain; Nature Communications finds 22 out of 26 aligned LLMs fully susceptible to ethical drift with 100% attack success rate; Cyberspace Administration mandates AI-generated content labeling in short videos, with 6 platforms removing 37,000违规 clips and relabeling over 600,000 items in the past month