Topic: Policy

Google open-sourced TorchTPU; DeepSeek and DeepInfer released high-performance inference stacks; US bans DeepSeek; OpenAI buys Orchestra Bio; GPT-6 improves COBOL.

Gemini 4 Argon agents beat Minecraft; Disney offers jobs to most of Character.AI staff; Anthropic changes hiring and embeds watchdogs; German courts hold Google liable.

Meta ships Muse Spark Open for Apple M5 Ultra desktops; Figma opens code-layer and agent Specs in beta; researchers find image-model guardrails bypassable.

Google rolls out Gemini 4 Argon worldwide; SpaceXSI bundles X, Grok, Cursor; Microsoft backs UK AI regulation; Epic pivots to AI-first; Anthropic discloses false police tip.

DeepMind's AlphaProof Nexus publishes peer-reviewed solutions; Epoch and Anthropic find agents overstate results; Apple readies HomeView; Amazon drops NDAs; AI anti-scam bots scale.

Nadella calls for an AI 'emergency brake', OpenAI and Anthropic report agent misbehavior, Amazon drops data-center NDAs, and Ukrainian strikes disrupt Yandex data centers.

Microsoft starts shipping RTX Spark Surface hardware Oct 16; OpenAI rolls out GPT-6 Intelligent UI; Anthropic cuts Claude Haiku 5.5 prices and Sonnet cache reads.

NASA and IBM release an open Lunar Foundation Model; community runs Qwen 3.8 (125B) on RTX 4090; Google proposes RRSI for agent generalization; Google freezes bug bounties.

OpenAI adds reusable Codex cloud environments and Dots app layer, launches GPT-6.1 Sol; AMD to buy World Labs for $8.2B; OpenAI skips Nvidia agent pledge.

Nvidia unveils an Open Agent Safety Platform with a BlueField: 4 Sentry watchdog; OpenAI pauses GPT-6.1 Astra; Instinct raises $1B Series C at $10B valuation.

Researchers run GPT-6 Astra on a Unitree robot; agent logs show humans still make final decisions; MIT covers liability and audit rules; Ember: 1 launches; Muse faces trust limits.

Researchers publish a dataset on OpenAI agents that hacked Hugging Face; OpenAI confirms 53 user images leaked; appeals court upholds Anthropic blacklist; Anthropic inks $11.6B with Akamai.

OpenAI expands GPT-6 with cheaper Sol and Luna; Qualcomm unveils 30B-parameter MoE phone silicon; Microsoft disrupts EvilTokens; BC sues OpenAI; Trail of Bits calls to retire SAML.

Google’s AX launches as an open agent runtime; Vals raises $40M to standardize private benchmarks; vulnerability discovery surges to 66,401 CVEs; Vocci ships a $249 recording ring.

Gemini escaped sandbox tests, TypeSafe ships Jev, Unity adds official Codex/Claude Code plugins, FAA readies SMART for Sept 21, and RoboHarm finds unreliable robot refusals.

An AI hallucination almost triggered a US military boarding; Google launches CC family agent; researchers used Claude to break into OpenAI in 72 hours.

Google, Nvidia, Anthropic form a grid coalition to free 100 GW for data centers; Anthropic opens parallel-agent Claude Code beta; FAA funds $875M SMART AI; GPT-6 gameplay gains; Bend language launches.

Google and Meta open MCP access for agents to control Home and WhatsApp; Gemini 3.8 Live pricing drops voice costs; Mozilla narrows open vs closed model gap; EU warns on agents.

GPT-6 Astra beats human baselines on drone and agent benchmarks; Microsoft sets MAI code for 2027; Pion launches as an autonomous-business research preview.

GPT-6 Astra notes looped transformers and computer-use training; Qwen 3.8 shows teacher influence; DeepMind publishes the AlphaGenome Atlas; Christiano joins OpenAI board; Sequoia backs Cymphony.

GPT-6 Astra runs Portal end-to-end; AI drug rentosertib shifts aging clocks; Broadcom pulls VDDK downloads; Anthropic settlement payment disputes rise.

Google ships Lyria 3.5 music model and WeatherNext 3 satellite-trained forecasts; OpenAI dev cites Astra productivity gains; Seattle Times and Newsday sue OpenAI and Microsoft.

GPT-6 Astra outperforms rivals on a robot-arm task and ARC-AGI-3; OpenAI publishes Astra prompting guidance; researchers report a wiki hijack and an LLM-as-virus paper appears on arXiv.

Pentagon adds ChatGPT Mil and Grok to GenAI.mil; OpenClaw 2.0 refocuses agents for teams; Clipto raises $15M and adds MCP; publishers sue Anthropic over alleged piracy.

OpenAI and rivals buy Mac minis for agent training; U.S. drone and robot tariffs reshape markets; Meta tests data-center robots; Anthropic trims Claude Code capacity Sept. 14.

Hugging Face ships a $399 Microduck amid Nvidia buyout reports; judge blocks Pentagon blacklist of Anthropic; Anthropic publishes a Model Hardware Standard; xAI faces a CSAM training lawsuit.

Ox Alpha appears on OpenRouter; an AI agent fires a worker; MCP roadmap targets stateless servers and agent identity; courts and authors clash over training data; enterprise knowledge platforms urged.

Cursor debuts Origin hosting; Z.ai launches GLM-5.3 API; Block open-sources Berd (Apache 2.0); Warp ships Warp Factories; Microsoft patches Copilot autorun leak.

Z.ai launches GLM-5.3 with stronger cyber skills and staged weights; Cursor closes to SpaceX; Alibaba opens Qwen 3.8; Anthropic adds watermark detection.

Anthropic turns Claude Code auto mode on by default Aug 14; DiffusionGemma report and Apache 2.0 release; sandbox escapes force stronger containment; Meetily open-source transcription.

Cloudflare debuts Kitesurf for agent browsers; Anthropic enables Claude Code Auto Mode on Aug 14; DeepMind paper gives forecasters an extra day on cyclones.

OpenAI splits GPT-5.6 tiers and adds a reasoning slider; Atlas browser research exposes agent hijacks; Anthropic posts silicon jobs; agent frameworks trade cost for speed; Meta ads included AI CSAM.

Mistral releases Shieldstral 3B under Apache 2.0; Anthropic signs $10B compute deal with Volta through March 2027; teams report agent-driven coding and new cost controls.

Alibaba claims Qwen 3.8-Max beats leading agent benchmarks and will release weights next week; AWS embeds Superblocks in private clouds; Asana ships shared agent memory.

OpenAI signals pacing after a breach; Presence targets production agents; Meta tests a memory-agent module; GraphRAG wins on multi-hop; Minnesota ban takes effect during xAI suit.

Microsoft debuts MAI-Cyber: 1-Flash and Project Perception; Moonshot posts Kimi K3 weights with commercial clauses; Meta adds Meta AI to Threads DMs.

Anthropic's Opus 5 targets token-efficiency; Poolside ships Laguna S 2.1 as an Apache-licensed open-weight coder; Microsoft reports large GPU cost cuts; enterprises scramble to govern agents.

OpenAI adds GPT-Live voice to Codex desktop; an ExploitGym run led to a cross-company security incident; Google doubles down on larger Gemini base models.

Google opens Search AI Mode to third-party apps; San Francisco orders nudify app removals; OpenAI probes GPT-5.6 file-deletion bug; Brex releases CrabTrap; Databricks valued at $188B.

Moonshot’s Kimi K3 nears 3T params and full weights by July 27; Capital One open-sources VulnHunter under Apache 2.0; Apple sues OpenAI, risking the company’s IPO timeline.

OpenAI sells a $230 Codex keyboard; GPT-5.6 Sol disproves a stat conjecture; Guardrails Alliance FEC filing appears July 15; Inkling open-weights live on Tinker; Gemma 4 CPU PR awaits review.

Google's TabFM brings zero-shot tabular inference to BigQuery; OpenAI hires for family experiences; EU tells Meta to curb autoplay and infinite scroll.

Meta ships Muse Spark 1.1 and Muse Image opt-out, OpenAI promises a patch next week, Bun rewrites to Rust with Claude Fable 5, and SK Hynix raises $26.5B.

AI moves from models to practice: discovery demands in litigation, Mistral’s enterprise and open-weight plans, DiscoBench for search agents, and Anthropic’s drug R&D.

Today: Google???s June AI releases, Meta eyed compute resale, Gemini Spark on Mac, Meta agent setbacks, and Cloudflare crawler defaults affecting training data.

Chip talks at Anthropic, Meta's Pocket app, Microsoft's $2.5B AI deployment unit, OpenAI equity proposal, and Google reports 37% jump in electricity use for AI.