AI News: May 22, 2026

1. Hark raises $700M Series A at $6B valuation for universal AI interface

Funding. Brett Adcock’s stealth AI lab Hark closed a $700M Series A led by Parkway Venture Capital at a $6B post-money valuation, with Nvidia, AMD Ventures, Intel Capital, Qualcomm Ventures, Salesforce Ventures, ARK Invest and Brookfield participating. The company plans to ship its first multimodal foundation models this summer and is developing dedicated personal-assistant hardware on a newly secured Nvidia B200 cluster. Source

2. Spotify and Universal Music sign landmark license for fan-made AI covers

Music. Spotify and Universal Music Group announced licensing agreements covering both recorded music and publishing to let Premium subscribers create AI-generated covers and remixes of catalog songs, with participating artists earning a revenue share on top of existing royalties. The deal puts Spotify in direct competition with Suno and Udio and is the first time the streaming platform has authorized user-driven AI music generation on its surface. Source

3. Trump delays AI security executive order

Policy. President Trump postponed signing an executive order that would have required federal security reviews of frontier AI models before public release, saying he did not want to slow US labs racing China. The pause leaves national-security testing of advanced models on a voluntary footing despite recent recommendations from US Cyber Command and AISI staff. Source

4. EU agrees to delay high-risk AI Act rules by up to 16 months

Policy. The Council and Parliament reached a political agreement on an “AI Act Omnibus” that pushes back enforcement of high-risk system obligations by up to 16 months and delays national regulatory sandboxes to August 2027, while shortening the transparency grace period for AI-generated content to three months. The package also extends prohibitions to “nudifier” applications generating non-consensual intimate imagery, effective December 2, 2026. Source

5. Connecticut SB5 becomes one of the strictest US state AI laws

Policy. Connecticut’s bipartisan SB5 passed on May 1 and was unpacked in detail this week, adding requirements for AI companion disclosures, synthetic-media watermarking and impact assessments for automated employment decisions. The law’s staggered effective dates begin in 2026 and overlap with the EU AI Act timeline, complicating compliance for cross-border vendors. Source

6. US Cyber Command stands up task force to deploy AI on classified networks

Government. US Cyber Command launched a dedicated task force to bring frontier AI models onto top-secret Pentagon and NSA networks, citing rising adversary use of automated vulnerability discovery. The unit will run accredited deployments behind air-gapped enclaves and is recruiting from major labs to build classified inference stacks. Source

7. 100-page UIUC/Meta/Stanford survey identifies three layers of agent harnesses

Research. A joint academic survey from UIUC, Meta and Stanford categorizes AI agent stacks into three harness layers (interface, mechanism, scaling), arguing that the majority of agent failures stem from infrastructure rather than model capability. The paper proposes a shared taxonomy for evaluating harnesses and includes reproducible audits across LangGraph, AutoGen and OpenHands. Source

8. Spotify launches NotebookLM-style desktop app for personalized podcasts

Tools. Spotify shipped a desktop research-preview app in 20+ markets that lets users feed documents and URLs into an AI pipeline that produces personalized daily or weekly podcast briefings, directly targeting Google’s NotebookLM. The release pairs with an ElevenLabs-powered audiobook creation tool that lets authors produce narrated editions without exclusive publishing deals. Source

9. The Path raises round for clinically benchmarked AI therapy

Healthcare. The Path, a mental-health startup founded by Tony Robbins and former Calm executives, launched with a custom AI model it says scores materially higher than consumer chatbots on standardized suicide-safety and crisis-routing benchmarks. The company is positioning around regulated therapy use cases rather than open-ended companionship. Source

10. Chatbot Arena update: Anthropic sweeps top four slots

Benchmarks. The latest LMSYS Chatbot Arena text leaderboard shows Anthropic occupying the top four positions with claude-opus-4-6-thinking (1502), claude-opus-4-7-thinking (1500), claude-opus-4-6 (1498) and claude-opus-4-7 (1492), followed by Meta’s muse-spark and Google’s gemini-3.1-pro-preview. The thinking variants outscore their non-thinking counterparts in both Opus generations, reinforcing reasoning-tuned models as the new arena ceiling. Source