AI INTELLIGENCE / DAILY SIGNAL
0011011011
2026-05-12 / AI NEWS

AI News Intelligence Digest — May 12, 2026

Altman Takes the Stand: "Are You Completely Trustworthy?" The most anticipated testimony of the Musk v. Altman trial arrived today. Musk's attorney Steven Molo opened cross-examination with that blunt question, then gril...

BRIEFING STRUCTURE21 SIGNALS

Inside this digest

  • Top Stories
  • Model Updates
  • Open Source
  • Industry Moves
  • Safety & Policy
  • Worth Watching
← Back to archive

Top Stories

  • Altman Takes the Stand: "Are You Completely Trustworthy?" The most anticipated testimony of the Musk v. Altman trial arrived today. Musk's attorney Steven Molo opened cross-examination with that blunt question, then grilled Altman on accusations of dishonesty from former board members and executives. Altman's own bombshell: he described a "hair-raising" 2017 moment when Musk suggested OpenAI should "pass to my children" if Musk died while controlling a proposed for-profit subsidiary. Altman said this confirmed his fear that Musk wanted personal control over AGI — the exact thing OpenAI was founded to prevent. CNN also reported Altman testified there was a "morale boost" when Musk left OpenAI. Testimony continues Wednesday; closing arguments expected Thursday. Developing — Week 3, now in climactic final testimony phase. (CNN; Bloomberg; TechCrunch; CNBC)

  • Google Detects First AI-Generated Zero-Day Exploit in the Wild. Published Sunday, confirmed today by Axios. Google's Threat Intelligence Group caught a cybercrime group using an AI model called "OpenClaw" to autonomously discover and build a zero-day exploit — a Python script that bypasses two-factor authentication on a web admin tool. Google determined "with high confidence" the exploit was LLM-generated based on hallucinated CVSS scores, educational docstrings, and structured code patterns characteristic of LLM training data. The group planned a mass exploitation event; Google and the affected vendor patched before damage spread. Separately, PRC and DPRK-linked groups are actively using AI for vulnerability research. This is no longer theoretical — AI-assisted offensive hacking is here. New story. (Google Cloud Blog; Bloomberg; The Hacker News; CNBC; Axios)

  • Google Launches "Googlebook" — AI-Native Laptops Built Around Gemini. Announced today at The Android Show: I/O Edition. Google is entering the premium laptop market with Googlebook, designed from scratch around Gemini Intelligence. Key feature: "Magic Pointer" — wiggle your cursor and Gemini surfaces contextual suggestions based on what's on your screen. Custom AI widgets, Android app compatibility, and phone-file access built in. Hardware partners: Acer, ASUS, Dell, HP, Lenovo. Ships this fall. Meanwhile, Gemini Intelligence is becoming the base layer of Android itself — not just an app, but the OS intelligence substrate. Gemini 4 expected at the main I/O keynote May 19. New story. (Google Blog; TechCrunch; Axios; Tom's Guide)

  • EU Reaches AI Act Omnibus Deal at 4:30 AM — Bans Nudifiers, Delays High-Risk Rules. The May 7 trilogue produced the most significant amendment to the EU AI Act since passage. Key changes: high-risk AI system deadlines pushed to December 2027 (standalone) and August 2028 (embedded products); new prohibition on AI-generated non-consensual intimate imagery and CSAM; watermarking requirements delayed to December 2026; regulatory sandbox established at EU level. The original August 2, 2026 full-applicability deadline stands for everything else. Formal adoption expected on accelerated timeline. New story. (EU Council; TechPolicy.Press; Bird & Bird)

Model Updates

  • Gemini 4 expected at Google I/O (May 19). Pre-I/O leaks suggest it scores 84.6% on ARC-AGI-2 (vs 68.8% for Claude Opus 4.6 and 77.1% for Gemini 3.1 Pro). If confirmed, this would be a significant reasoning jump. (Android Authority; BusinessToday; Hacker News)
  • No new frontier model drops today. Current standings: Claude Opus 4.7 (coding/finance), GPT-5.5 (agentic/default ChatGPT), DeepSeek V4 Pro (open-source leader), Gemini 3.1 Ultra (multimodal). (LLM Stats)
  • GLM-4.7 (Zhipu AI) continues gaining attention — 1.2% hallucination rate, trained entirely on Huawei Ascend silicon, $0.11/M input tokens vs. Claude Opus at $15. The NVIDIA-free frontier keeps expanding. (LLM Stats; The Register)

Open Source

  • Gemma 4 (Google) — new open-weight model family announced ahead of I/O. Details expected May 19. (Google Blog)
  • NVIDIA Ising — world's first open-source AI models purpose-built for quantum computing error correction. 2.5x faster and 3x more accurate decoding vs. traditional approaches. (NVIDIA Newsroom)
  • NVIDIA Nemotron 3 — Nano/Super/Ultra family optimized for agentic AI. Claimed most efficient open models at their accuracy levels. (NVIDIA Newsroom)

Industry Moves

  • Google investing up to $40B in Anthropic ($10B now at $350B valuation, $30B conditional on performance milestones). Anthropic has committed ~$200B to Google Cloud TPUs over five years (3.5-5GW capacity). Revenue now $30B ARR, up from $9B end of 2025. (TechCrunch; CNBC; Anthropic)
  • NVIDIA passes $40B in AI equity bets for 2026 — led by $30B OpenAI investment, plus Corning ($3.2B) and IREN ($2.1B). NVIDIA is becoming an infrastructure investor, not just a chip vendor. Developing — from last week. (CNBC; The AI Insider)
  • Q1 2026 venture funding: $300B — 80% ($242B) went to AI. Four of the five largest VC rounds ever closed in Q1: OpenAI ($122B), Anthropic ($30B), xAI ($20B), Waymo ($16B). (Crunchbase)
  • Pentagon struck deals with 8 Big Tech companies for AI — notably shunning Anthropic despite its cybersecurity credentials. (CNN)

Safety & Policy

  • AI-assisted zero-day exploits are now real (see Top Stories). The Palisade Research self-replication findings from last week (81% success rate) combined with today's Google disclosure paint a clear picture: offensive AI capabilities are advancing faster than defensive preparations. (Google TIG; Palisade Research)
  • EU AI Omnibus (see Top Stories) — the nudifier ban and CSAM provisions represent the first hard prohibition of specific generative AI outputs in major legislation. (EU Council)
  • White House FDA-Style AI Vetting EO still expected within days — no new updates today, but the Google zero-day report adds urgency. Developing — from last week. (The Hill; Bloomberg)
  • Global AI adoption: 17.8% of working-age population now using AI (up 1.5pp in Q1 2026). 72 countries have proposed 1,000+ AI policy initiatives. (Microsoft; Mind Foundry)

Worth Watching