JadidJust-in AI Developments, Incidents & Deadlines
May 2026
SMTWTFS12345678910111213141516171819202122232425262728293031

AI news, Tuesday, May 19, 2026

9 stories

TLDR

  • Google I/O 2026 keynote today, 10am PT / 1pm ET (May 19) — Sundar Pichai opens at Shoreline; expected: new Gemini model (Gemini 3.5 Pro "Cappuccino" leaked; Gemini Omni unified text/image/video model widely reported), Android XR glasses preview with Samsung/XREAL, Android 17, and Workspace/Search Gemini integrations.
  • Anthropic dethrones OpenAI atop CNBC Disruptor 50 (May 19) — Annualized revenue jumped from ~$9B (EOY 2025) to ~$30B by Q1 2026 (Amodei: "80x growth in Q1"); OpenAI takes #2.
  • Dust closes $40M Series B for "multiplayer" enterprise AI (May 18) — Sequoia/Abstract-led; 3,000+ orgs, 300k+ agents deployed, 240% NRR with zero churn. Signal that the enterprise agent layer is consolidating around shared-context platforms, not single-user copilots.
  • Recursive Superintelligence out of stealth at $4.65B (May 13) — Richard Socher + Tim Rocktäschel raised $650M from GV, Greycroft, Nvidia, AMD to build self-improving AI; products promised "in quarters, not years."
  • CVE-2026-27740: Discourse AI plugin XSS via LLMLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ output (May 2026) — LLM-generated content was interpolated unescaped into templates across LLM Triage, Personas, and Agents plugins. Patched in 2026.3.0-latest.1 / 2026.2.1 / 2026.1.2.

Action Items

Things your engineering team should look at or do this week.

  • Block off the next 24 hours for Google I/O API fallout — Most pricing/model-string changes drop within hours of the keynote (10am PT / 1pm ET today). Designate one engineer to watch the AI Studio/Vertex changelog, then track which of your Gemini deployments need updated model IDs, region availability changes, or rate-limit re-baselining. Engadget and Android Central are running live blogs if you're not watching the stream. (Engadget live blog | Android Central live blog)
  • Patch Discourse if you run one — CVE-2026-27740 hits any team running Discourse with LLM Triage/Personas/Agents plugins (common for community/support forums); LLM output rendered unescaped means an attacker can inject HTML/JS via crafted prompts that the model then echoes. Upgrade to 2026.3.0-latest.1, 2026.2.1, or 2026.1.2. Reasonable second action: audit your own LLM-output-to-template paths for the same class of bug. (SentinelOne advisory)
  • If you use LiteLLM proxy, confirm you're on ≥1.83.10-stable — CVE-2026-42208 (CVSSCommon Vulnerability Scoring SystemA standard 0 to 10 severity score: 0.1–3.9 low, 4.0–6.9 medium, 7.0–8.9 high, 9.0–10 critical.FIRST ↗ 9.3) was added to CISACybersecurity and Infrastructure Security AgencyThe U.S. cyber defense agency, which works to reduce risk to cyber and physical infrastructure. It runs KEV.CISA ↗'s KEVKnown Exploited VulnerabilitiesCISA's list of vulnerabilities with reliable evidence of active exploitation and a clear fix, used to set patching priority.CISA ↗ catalog May 8 after active exploitation within 36 hours of disclosure. Federal civilian deadline was May 11; if you missed your own patching window, treat it as priority-1 today. (Sysdig writeup | LiteLLM advisory)

Models & Tools

  • Google I/O 2026 keynote: new Gemini family expected (May 19, 10am PT) — Pre-event signals: a model codenamed "Cappuccino" surfaced in AI Studio/LM Arena (probable Gemini 3.5 Pro), Gemini 3.2 Flash quietly appeared in the iOS app and AI Studio on May 5, and internal docs leaked the "Gemini Omni" name for a unified text/image/video/audio architecture. Pichai-led keynote also expected to cover Android XR glasses (Samsung "Jinju," XREAL Project Aura), Android 17, and Search/Workspace agent surfaces. Treat anything below as preview-only until the official changelog publishes. (Tom's Guide preview | TheNextWeb preview)

  • Hugging Face wires open-source LLMs into GitHub Copilot Chat (mid-May) [Open Source] — A new extension lets developers route Copilot Chat requests inside VS Code to open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ models hosted on Hugging Face InferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗. Practical use case: cheap or air-gapped models for non-sensitive tasks, premium frontier for the rest, all from the same Copilot surface. (DevOps.com writeup)

  • GFT ships AI inspection-to-action robotics for auto OEMs (May 18) [Robotics/IoT] — GFT Technologies launched AI-powered robotic arms for automotive assembly lines that detect defective parts via computer vision and physically remove them — closing the loop between AI inspection and physical action that most "smart factory" pilots still leave to humans. Worth tracking if your org has any physical-world AI ambitions; this is the pattern most industrial AI roadmaps converge on. (Robotics & Automation News)

Security

  • CVE-2026-27740: Discourse AI plugin stored-XSS via LLM output (May 2026) [Production] — In Discourse's LLM Triage, Personas, and Agents plugins, model-generated content was interpolated into templates without HTML escaping. Crafted prompts (or upstream poisoning) could embed JavaScript that executes in the browser of any user viewing the AI output. Fixed by applying ERB::Util.html_escape across all affected code paths in 2026.3.0-latest.1, 2026.2.1, and 2026.1.2. The pattern — trusting LLM output as "data" when downstream it's interpreted as "code" — is the kind your own pipelines probably need to audit for too. (SentinelOne advisory)

Industry & Business

  • CNBC Disruptor 50 2026: Anthropic #1, OpenAI #2 (May 19) — First time OpenAI has been displaced at the top of the list. CNBC cites Anthropic's $30B annualized run-rate (vs. ~$9B at end of 2025), Claude Code as a major revenue driver, and Amodei's claim of 80x revenue growth in Q1. 43 of the 50 list companies say AI is essential to their business model. Total funding across the Disruptors hit $337B, up from $127B in 2025. (CNBC: Why Anthropic is #1 | CNBC: Full list)

  • Dust raises $40M Series B for "multiplayer AI" in the enterprise (May 18) [Agents] — Round led by Abstract and Sequoia, with Snowflake and Datadog participating. Pitch: enterprise AI agents shouldn't live in 1:1 chat windows — they should sit in a shared collaborative surface with persistent context across people, tasks, and tools. Metrics it's pitching on: 3,000+ orgs, 300k+ deployed agents, 90%+ MAU and 70%+ WAU, 240% NRR with zero churn in 2025. Useful data point if you're sizing "is enterprise agent adoption real yet." (SiliconANGLE | Tech.eu)

  • Recursive Superintelligence emerges from stealth: $650M at $4.65B (May 13) — London-based; Richard Socher (ex-Salesforce chief scientist) and Tim Rocktäschel (UCL/ex-DeepMind) leading. Round co-led by GV and Greycroft, with Nvidia and AMD in. Mission is recursively self-improving AI — models that analyze and rewrite themselves without human feedback. Socher says product launches in "quarters, not years." Note this lands two weeks after Ineffable Intelligence's $1.1B seed (also reinforcement-learning-bet, also ex-DeepMind founder); investors are clearly funding multiple shots at "post-RLHF" approaches. (Tech.eu | TechCrunch analysis)

AI in Production

  • Cursor 3.3 — parallel agent execution, in-IDE PR review (May 7) [Agents] — Cursor 3.3 added a "Build in Parallel" mode that splits a plan into independent async subagents, plus a Reviews tab that lets you take a PR from creation to merge inside the editor and a quick action to split a chat-driven change into multiple logical PRs. Mostly relevant if you've been waiting for Cursor's plan/agent UX to mature enough to justify a real team rollout — this is the release where parallelism becomes default. (Cursor changelog)

Policy & Safety

  • arXiv to ban authors who submit obvious LLM-generated papers (May 2026) — Computer-science section chair Thomas Dietterich confirmed that authors caught submitting unchecked LLM-generated content will be barred from submitting for a year. Framed as an integrity measure against a growing flood of AI-generated submissions. Worth noting if your team publishes preprints — even legitimate LLM-assisted writing now needs a defensible audit trail of human verification. (404 Media)

Type to search · ↑↓ to move · enter to open