OpenAI and Dell partner to bring Codex to on-premises enterprise

OpenAI and Dell Technologies announced a partnership to deploy Codex in hybrid and on-premises enterprise environments, extending OpenAI's agentic coding assistant to customers who cannot or will not run workloads in public cloud. The deal targets the large share of enterprise IT that remains behind the firewall and pairs Codex with Dell's infrastructure distribution reach — a channel OpenAI has not previously had direct access to.

Anthropic ships Claude for Microsoft 365, Excel and Word go GA

Claude for Excel, PowerPoint, and Word reached general availability for paid plans, with Claude for Outlook entering public beta. The integrations keep conversation context as Claude moves between apps, and the rollout makes Claude the first major AI assistant besides Microsoft's Copilot to hold native, cross-app context inside the M365 suite.

Meta launches Incognito Chat with Meta AI on WhatsApp

Meta rolled out Incognito Chat on WhatsApp and the Meta AI app, offering temporary AI conversations that Meta says even its own engineers cannot read, using on-device trusted execution environments. The launch arrived five days after Meta quietly removed end-to-end encryption from Instagram DMs — a sequence that drew immediate criticism from the EFF as contradictory on privacy.

Anthropic adds MCP tunnels and self-hosted sandboxes to Managed Agents

Anthropic shipped two new security features for Claude Managed Agents: MCP tunnels, which route agent tool calls through a private network without exposing internal services to the public internet, and self-hosted sandboxes, which let enterprises run tool execution inside their own infrastructure rather than Anthropic's. The update directly addresses the enterprise objection that cloud-hosted agent runtimes require sensitive internal systems to be externally reachable.

Gemini 3.5 Flash Beats Gemini 3.1 Pro on Coding and Agentic Benchmarks at 4x Speed

Announced at Google I/O on May 19, Gemini 3.5 Flash scores 76.2% on Terminal-Bench 2.1 coding evaluation and 1656 on GDPval-AA real-world agentic benchmark—surpassing its predecessor Gemini 3.1 Pro on both coding and multimodal tasks while running 4x faster on output tokens at roughly half the cost. The model ships today as the default across the Gemini app, Google Search AI Mode, and the Gemini API. Caveats: benchmarks are self-reported by Google, and practitioners on Latent Space note weak scores on TerminalBench-Hard and mediocre MRCR/ARC-AGI-2, with some flagging that GPT-5.5-medium may still be faster or smarter end-to-end.

Google Antigravity 2.0 Demos 93-Agent OS Build in 12 Hours for Under $1K

Also at I/O, Google's Antigravity 2.0 agentic coding system completed the core framework of an operating system in roughly 12 hours by spinning up 93 parallel sub-agents, issuing 15,000+ model requests, and consuming 2.6 billion tokens—all for under $1,000 in API credits. The prior state of the art for comparable multi-agent scaffolding required dedicated cluster time and manual orchestration; Antigravity wraps the entire pipeline into a single harness now available via the Gemini API and used to power Gemini Spark. The demo is Google-run and not yet independently replicated.

Meta FAIR's AIRA agents autonomously design neural architectures, beat human baselines at 1B scale

Researchers at FAIR (Meta) built LLM agent frameworks that autonomously designed new neural architectures — dubbed AIRAformers and AIRAhybrids — outperforming established human-designed models including Llama 3.2 and Nemotron-2 at the 1B-parameter scale. The agents also optimized attention mechanisms and training scripts competitively. If the result holds up to scrutiny, it marks a concrete step toward automating architecture search at production model sizes.

PhysBrain 1.0 grounds embodied AI in egocentric video for real-world robot tasks

PhysBrain 1.0 integrates physical commonsense derived from human egocentric video into a vision-language-action model, improving physically grounded reasoning across VLM benchmarks, VLA simulations, and real-world robot tasks. The technical report shows higher task success rates than prior VLA baselines in manipulation scenarios requiring contact-rich reasoning. The approach sidesteps the need for large robot-specific datasets by transferring physical priors from human video.

Trump executive order to give government early frontier-model access

A draft executive order under active White House review would give the U.S. government first access to new frontier AI models before public release, stopping short of blocking deployment but establishing a formal pre-clearance process. The order would create a working group of tech executives and federal officials to design the oversight framework, and the White House cyber office is separately developing an AI security standard requiring Pentagon safety testing before government deployment. The move marks a reversal for an administration that opened its term by rescinding Biden-era pre-deployment review requirements.

Colorado governor signs SB 189, replacing landmark AI discrimination law

Governor Jared Polis signed SB 189 into law on May 14, 2026, repealing and replacing Colorado's original AI Act with a revised framework now covering 'automated decision-making technology' (ADMT). The new law restructures liability between developers and deployers, imposes dual notice requirements at the point of interaction and after adverse decisions, and takes effect June 20, 2026 — making Colorado the first state to enact, then substantially overhaul, a comprehensive AI governance statute before it ever took effect.