OpenAI spins out Deployment Company, acquires Tomoro for enterprise push

OpenAI launched the OpenAI Deployment Company, a standalone business unit seeded with $4 billion to embed Forward Deployed Engineers inside enterprise clients. To staff it immediately, OpenAI acquired Tomoro, a firm with real-world AI deployments at Tesco, Virgin Atlantic, and Supercell. The move signals OpenAI shifting from a model vendor to an end-to-end systems integrator, betting that the next enterprise revenue wave comes from deployment depth, not raw model capability.

Anthropic ships Claude Managed Agents and Claude for Small Business

Anthropic's May developer conference yielded two platform expansions: Claude Managed Agents—adding dreaming, multiagent orchestration, outcomes, and webhooks to the API—and Claude for Small Business, a package launched May 13 that embeds Claude inside the tools small operators already use, built on the Claude Cowork agent layer. The small-business push is a direct attempt to close an adoption gap: small businesses make up 44% of US GDP but have lagged enterprise AI uptake.

Google ships Gemini Intelligence on Android at I/O 2026

Announced May 12 at Google I/O, Gemini Intelligence brings proactive AI features to Android—automating multi-step tasks, summarizing web content, filling complex forms, and letting users build custom widgets via natural language. A Rambler feature converts spoken messages into polished text. The rollout starts this summer on select Samsung and Pixel devices, marking Google's most aggressive Gemini integration into the core Android OS to date.

xAI launches Grok Build, its first agentic coding CLI

Grok Build launched May 15 in early beta, gated to SuperGrok Heavy subscribers at $300/month. The terminal-native agent can spawn up to eight concurrent sub-agents, runs on Grok 4.3 with a 2-million-token context window, and integrates with VS Code. xAI is explicitly positioning it against Claude Code and OpenAI Codex—Elon Musk has previously admitted xAI has fallen behind competitors on coding, and this is the direct response.

ClawBench Puts Web Agents on Live Production Sites—Top Score Is 33%

UBC and the Vector Institute released ClawBench, an evaluation framework of 153 tasks across 144 live production websites covering purchases, bookings, and job applications—the first major agent benchmark to run on real sites rather than sandboxes. The best frontier score belongs to Claude Sonnet 4.6 at 33.3%, a number that clarifies how far agentic browser capability still has to travel before it can be called reliable. The benchmark captures five layers of behavioral data per run and scores with an agentic evaluator that produces step-level diagnostics, making it harder to game than sandbox alternatives.

Frontier AI Cuts Palo Alto Networks' Pen-Test Year to Three Weeks

Palo Alto Networks published results from scanning all 130+ of its products using frontier AI models—including Anthropic's Claude Mythos and OpenAI's GPT-5.5-Cyber—as part of its May 'Patch Wednesday' disclosure cycle, marking the first time the majority of findings came from AI-assisted scanning rather than human testers. Their reported finding: three weeks of model-assisted analysis matched the coverage of a full year of manual penetration testing. The same testing showed AI-assisted attack cycles compressing time from initial access to exfiltration to as little as 25 minutes, setting a new practical bar for mean-time-to-respond requirements.

Anthropic open-sources Bloom, an agentic behavioral evaluation framework

Bloom is an open-source agentic framework that takes a researcher-specified behavior and automatically generates scenarios to quantify its frequency and severity across frontier models — no hand-labeled transcripts needed. In testing, Bloom reproduced the same model-ranking results as manual system card evaluations and uncovered a new finding: increased reasoning effort reduces self-preferential bias in Claude Sonnet 4. The tool addresses a core bottleneck in safety evaluation pipelines, where manual transcript labeling doesn't scale to the pace of model releases.

AMD's MI350P PCIe card brings 4 TB/s HBM3E to standard enterprise servers

AMD introduced the Instinct MI350P, a dual-slot air-cooled PCIe accelerator designed for enterprises that can't swap out existing server infrastructure for liquid-cooled GPU clusters. The card delivers an estimated 2,299 TFLOPS (peaking near 4,600 TFLOPS at lower precision) with 144 GB of HBM3E memory and 4 TB/s bandwidth, supporting MXFP4, MXFP6, FP8, and INT8 formats. It targets inference workloads — including agentic AI and RAG pipelines — and supports up to eight cards per air-cooled system, giving enterprises an incremental on-prem inference path without full GPU platform redesign.

Colorado governor signals SB 189 signing, replacing landmark AI discrimination law

Colorado Governor Jared Polis confirmed he will sign Senate Bill 189, a scaled-back AI disclosure bill that replaces SB 24-205—the country's first comprehensive state AI anti-discrimination statute—which was heading for a June 30 effective date but faced a federal court stay and an xAI lawsuit. The new bill, which passed 56-7, drops sweeping algorithmic-discrimination mandates and instead requires entities to notify individuals when AI contributes to adverse decisions on loans, hiring, or other consequential matters. The xAI v. Weiser litigation remains live: under a court-brokered agreement, Musk's company will file for a preliminary injunction within 28 days of finalized rulemaking under whichever law ultimately takes effect.

Connecticut omnibus AI bill clears legislature, awaits governor's signature

Connecticut's Senate Bill 5—one of the broadest state AI measures in the country—passed both legislative chambers on May 1, 2026 and now awaits the governor. The bill covers automated employment decisions, AI companion chatbots (banning romantic or manipulative interactions with minors), synthetic content disclosure, and a voluntary safe-harbor program administered by the Department of Consumer Protection; it also amends state anti-discrimination statutes to expressly cover AI-driven hiring processes. Connecticut becomes at least the third state after Colorado and California to advance a multi-domain AI governance framework in 2026, deepening the state-level patchwork that the Trump administration has so far failed to preempt through Congress.