AI News: May 19, 2026

1. Take It Down Act’s platform compliance deadline lands today, with FTC enforcement going live

Washington. The federal Take It Down Act’s notice-and-removal mandate becomes enforceable today, May 19, requiring covered platforms to take down nonconsensual intimate imagery, including AI-generated deepfakes, within 48 hours of a valid takedown request. The criminal prohibition on knowingly publishing or threatening to publish such imagery has been in force since the bill was signed a year ago, but today is the FTC’s first day of enforcement against platforms, with civil penalties of up to roughly $53,000 per violation treated as unfair or deceptive trade practices. FTC Chair Andrew Ferguson sent formal warning letters in advance of the deadline to Meta, Apple, Microsoft, TikTok, Reddit, Snapchat, and X, signalling the agency intends to test enforcement quickly against the largest user-generated content surfaces. Source

2. SandboxAQ wires its drug discovery models into Claude, betting on access over PhD-grade tooling

Drug discovery. SandboxAQ released a Claude integration that exposes its drug discovery and computational chemistry models directly to Claude users, with the explicit pitch that the bottleneck in pharma research is access to advanced computational tools rather than the algorithms themselves. The integration lets non-specialist researchers ask Claude to run SandboxAQ’s models for protein structure prediction, binding affinity estimation, and other computational chemistry tasks through natural-language requests instead of through a separate scientific computing stack. It is one of the more concrete examples this month of agentic-AI plumbing closing the gap between specialist scientific tooling and the general-purpose assistant surface. Source

3. Nous Research wires Codex CLI into Hermes Agent as a shell, billing through ChatGPT subs

Open source. Nous Research released Hermes v2026.5, which uses OpenAI’s Codex CLI as the runtime backend for the Hermes Agent. ChatGPT subscribers can now run Hermes-mediated sandboxed shell commands and structured patches through their existing ChatGPT subscription instead of buying separate API credits, with Hermes acting as the shell on top of Codex’s engine. The arrangement is in beta and ships with four agent-loop tools unavailable, which limits some end-to-end workflows, but it is an early concrete example of independent agent shells riding the major lab CLIs to bypass API-pricing barriers for hobbyist and prosumer users. Source

4. Havoc raises $100M Series A for all-domain collaborative autonomy

Funding. Havoc, an all-domain collaborative autonomy company, closed a $100 million Series A, bringing total capital raised to roughly $200 million since 2024. The round lands in the increasingly capitalized “AI for defense and autonomous systems” segment alongside Anduril, Shield AI, and Helsing, where the pitch is centered on coordinating heterogeneous unmanned platforms across air, ground, and sea through shared autonomy stacks. The deal sizing is notable for a Series A and reflects how aggressively defense-adjacent autonomy is being funded as US and allied governments push for production-scale unmanned systems. Source

5. GridCARE closes $64M Series A to industrialize “power acceleration” for AI buildouts

Funding. GridCARE, which positions itself as a pioneer of “Power Acceleration for AI,” closed a $64 million oversubscribed Series A led by Sutter Hill Ventures. The thesis is that the binding constraint on US data center expansion is no longer chips but grid interconnect and on-site power orchestration, and that AI-specific power infrastructure is now a distinct category with its own venture-scale opportunities. The round joins a small but visible wave of power-and-cooling startups funded against the AI capex cycle, including the substations and behind-the-meter generation segments that hyperscalers and neoclouds are now pre-purchasing years in advance. Source

6. Google I/O 2026 keynote opens today with Gemini, Android XR, and Aluminium OS on deck

Google. Google I/O 2026 kicks off at 10 a.m. PT at Shoreline Amphitheatre, with the two-day developer conference running May 19-20. The opening keynote is structured around four pillars Google has been pre-trailering for the last week: a new Gemini model, an agentic AI push, Android XR hardware, and the unveil of Aluminium OS as the Android-ChromeOS merger that will ship under the Googlebook laptop brand this fall. The conference is widely framed as Google’s biggest reset of its AI narrative since Bard, against a competitive backdrop where Anthropic’s Claude Mythos and OpenAI’s GPT-5.5 have set the recent frontier ceiling. Source Also covered in today’s Google roundup.

7. Jensen calls AI demand “utterly parabolic” as Dell’s Vera Rubin NVL72 PowerEdge ships for agentic inference

NVIDIA. Jensen Huang used his Dell Technologies World keynote on May 18 to call AI demand “parabolic, utterly parabolic,” and anchored the message around a new Dell PowerEdge generation built on the NVIDIA Vera Rubin NVL72 platform aimed at large-scale agentic inference at sharply lower cost-per-token. The flagship liquid-cooled PowerEdge XE9812 is the headline SKU on Vera Rubin NVL72, with XE9880L, XE9882L, and XE9885L filling out an HGX Rubin NVL8 tier for existing data center power envelopes, and the R9822 and M9822 debuting NVIDIA Vera CPU options for enterprise workloads. Huang named Lilly, Samsung, and Honeywell as enterprises already deploying Dell AI Factories at scale, framing this as the moment “AI factories” stop being a marketing pitch and start showing up as line items on enterprise capex plans. Source Also covered in today’s NVIDIA roundup.

8. Claude Code 2.1.144 brings /resume to background sessions and a worktree opt-out

Anthropic. Claude Code 2.1.144 lands the first version of /resume that can pick up background sessions, so a session started with claude --bg or from agent view now shows up alongside interactive sessions in the resume list and is labelled with bg to keep the two surfaces distinguishable. The same release adds worktree.bgIsolation: "none" for users who explicitly want background work to edit the working copy directly without going through EnterWorktree, an escape hatch for repos and workflows where the isolation that landed in earlier 2.1.x releases gets in the way. Smaller adds include elapsed-duration timestamps on background subagent completion notifications and a “last updated” column in the /plugin browse and /plugin discover panes so it is easier to tell stale entries apart from actively maintained ones. Source Also covered in today’s Anthropic roundup.