Daily Brief Archive
Explore our complete collection of AI insights. Every edition, every breakthrough, all in one place.
OpenAI acquires open-source evaluation and red-teaming platform Promptfoo to strengthen agent safety and compliance; NVIDIA partners with Thinking Machines Lab to deploy at least 1 gigawatt of Vera Rubin systems and invests, with deployment planned for early next year; Gemini launches multiple generation and retrieval features in Workspace, with Sheets achieving 70.48% success rate on SpreadsheetBench
Microsoft launches Copilot Cowork and announces Agent 365 will go commercial on May 1 at $15/user/month; Anthropic sues the U.S. Department of Defense over its 'supply chain risk' designation, triggering a six-month federal directive to discontinue Claude-related technologies; UK-based Nscale raises $2B in Series C funding, reaching a $14.6B valuation with participation from NVIDIA and others
OpenAI releases GPT-5.4 with 75% OSWorld-Verified success rate and 1 million token context support; OpenAI robotics lead Caitlin Kalinowski resigns over Pentagon collaboration deal; Security report reveals 42,089 exposed OpenClaw instances with CVSS 8.8 vulnerability
OpenAI launches Codex Security: scanning over 1.2 million commits in 30 days, identifying 792 critical vulnerabilities and assigning 14 CVEs; Google open-sources gws to unify Workspace APIs with MCP compatibility, gaining over 14,000 GitHub stars; audit reveals 45.83% of Shadow API endpoints fail model fingerprint verification, involving at least 116 papers relying on unreliable interfaces
Pentagon labels Anthropic a supply chain risk and requires contractors to stop using Claude; NVIDIA halts H200 production for China and shifts TSMC capacity to Vera Rubin; Clinejection discloses prompt injection combined with Actions cache poisoning leading to theft of NPM production release keys
OpenAI releases GPT-5.4 and opens access to Plus/Team/Pro and API users, reporting a 33% reduction in single-statement error rate compared to GPT-5.2; YuanLab releases and claims to have open-sourced Yuan 3.0 Ultra, a 1 trillion parameter MoE model with 68.8 billion activated parameters and 49% improvement in pretraining efficiency; IREN orders over 50,000 Nvidia B300 GPUs, expanding its AI cluster to approximately 150,000 GPUs with a 50% capacity increase
Nvidia says it may stop major follow-on investments after committing $30B to OpenAI and $10B to Anthropic; Microsoft reveals Phi-4-reasoning-vision-15B was trained on only about 200B tokens using a 'hybrid reasoning' approach; Stanford publishes Merlin, a 3D vision-language model for abdominal CT, in Nature and open-sources a dataset of 25,494 CT-report pairs
Google previews Gemini 3.1 Flash-Lite: $0.25 per million input tokens, first-token latency improved by 2.5x; Accenture reportedly acquiring Ookla for $1.2B and integrating over 250 million monthly network speed tests; OpenAI's internal data agent covers 600PB of data and serves approximately 4,000 employees
Galbot secures 2.5 billion yuan funding with National IC Fund Phase III's first investment in embodied intelligence; Xiaopeng's second-gen VLA features 22B parameters, 32x inference efficiency gain, rolling out late March; FireRed-OCR open-sources 2B model scoring 92.94% on OmniDocBench v1.5