JadidJust-in AI Developments, Incidents & Deadlines
June 2026
SMTWTFS123456789101112131415161718192021222324252627282930

AI news, Wednesday, June 17, 2026

11 stories

TLDR

  • NVIDIA dropped the Nemotron 3 open-model family (Jun 16) — open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ speech (ASR), multimodal RAGRetrieval-augmented generationFetching relevant documents at query time and giving them to the model, so its answer draws on them.Lewis et al. 2020 ↗, and safety models plus 10T open training tokensA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗; pitched as the strongest US open-weight stack Open Source
  • SearchLeak: a one-click M365 Copilot bug could exfiltrate emails, files, and MFA codes (Jun 15) — chained prompt injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ + HTML race + Bing SSRFServer-side request forgeryA bug where an attacker makes a server send requests for them, reaching internal services it shouldn't expose.OWASP ↗, max severity (CVE-2026-42824), now patched server-side Security
  • OpenAI shipped "Deployment Simulation" (Jun 16) — replays ~1.3M de-identified past conversations through a candidate model pre-release to estimate real-world failure rates Trending
  • ChatGPT's assistant market share fell below 50% for the first time (Jun 16) — ~46% vs Gemini ~28% and Claude ~10% by one measure Trending
  • AI took center stage at the G7 in France with Altman, Amodei, and Hassabis all present (Jun 15–17) — first time all three frontier-lab CEOs appeared before world leaders together Policy

Action Items

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

  • Patch any self-hosted LangGraph deployments now. A critical RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ chain in self-hosted LangGraph agents (SQL injection + unsafe deserializationAn app turning untrusted serialized data back into objects, which can let an attacker tamper with logic or run code.OWASP ↗, e.g. CVE-2025-67644 chained with CVE-2026-28277) was detailed ~Jun 12 and can lead to full server takeover. Upgrade to current releases and lock down checkpoint serialization. (Cyberpress writeup)
  • Evaluate NVIDIA Nemotron 3 for any on-prem or open-weight agent work. The new open releases cover real-time ASR, multimodal RAG, and safety/guardrail models — worth benchmarking against your current hosted endpoints before renewing closed-API spend for those workloads. (NVIDIA newsroom)

Models & Tools

  • NVIDIA Nemotron 3 — open-model family for agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ + physical AI (Jun 16) Open Source — NVIDIA released open-weight Nemotron 3 models (including leaderboard-topping speech/ASR, multimodal RAG, and safety models) alongside Cosmos, Alpamayo (AV), Isaac GR00T (robotics), and Clara (biomedical), plus large open datasets (10T language tokens, 500k robotics trajectories). Early adopters named include Salesforce, ServiceNow, CrowdStrike, Palantir, and Uber. (NVIDIA newsroom | NVIDIA blog)

  • OpenAI Deployment Simulation — pre-release behavior forecasting (Jun 16) Trending — A method that strips the assistant reply from ~1.3M de-identified historical conversations (GPT-5 Thinking → GPT-5.4, Aug 2025–Mar 2026), reruns the prompts through a candidate model, and grades the outputs to estimate live failure rates before launch (reported median multiplicative error ~1.5x). Now extended to agentic/tool-call coding scenarios. Useful prior art for teams building their own pre-deploy eval harnesses. (OpenAI)

  • Microsoft Work IQ hits GA; "Autopilot" agents push proactive automation (Jun 16) Agents — Work IQ APIs (organizational context/intent signals, generic MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ tools, semantic index) went GA Jun 16, underpinning Microsoft's new always-on "Autopilot" agent category. The first, Scout (announced Jun 2, built on the open-source OpenClaw framework), runs continuously across Teams/Outlook/OneDrive/SharePoint rather than waiting for prompts. (Microsoft 365 blog)

  • Google ships Android 17 with Gemini Omni and Lyria 3 (Jun 16) Customer AI — The June Pixel Drop brings Gemini Omni (prompt-based video creation/editing for Gemini Pro users), Lyria 3 (text/image-to-music), and on-device speech translation, plus Wear OS 7. Signals Google's continued push to bake generative media directly into the mobile OS. (Thurrott | How-To Geek)

Security

  • SearchLeak: one-click data exfiltration via M365 Copilot (Jun 15) Security Agents — Varonis Threat Labs disclosed a chain (CVE-2026-42824, max severity) combining parameter-to-prompt injection via a URL q parameter, an HTML rendering race condition, and a Bing SSRF — letting a single click on a legitimate microsoft.com link silently search a victim's mailbox and exfiltrate emails, files, and MFA codes. Microsoft has fully patched it server-side; no user action required, but a sharp reminder that AI assistants expand the phishing blast radius. (Varonis | The Hacker News)

  • LangGraph critical RCE chain in self-hosted agents (~Jun 12) Security Agents — Researchers detailed multiple vulnerabilities in self-hosted LangGraph (SQL injection + unsafe deserialization) that can be chained to remote code execution and full server takeover. A parallel LiteLLM proxy takeover chain (Obsidian Security, Jun 15) lets low-privilege users seize AI gateway servers — both underline that agent infrastructure is now an attack surface in its own right. (Cyberpress)

Industry & Business

  • Bezos backs CuspAI in a reported $400M round at a $2.6B valuation (Jun 17) Trending — Bezos Expeditions and Kleiner Perkins are backing the two-year-old Cambridge, UK startup applying generative AI to materials science (advisers include Hinton and LeCun). Term sheets signed, deal not yet closed; ~4x its Sept 2025 valuation — another big bet on AI for the physical world. (Silicon Republic / FT | Sifted)

  • World-models startup Odyssey raises $310M from Amazon and AMD (Jun 17) Infra — The Palo Alto lab building physics-trained world models (3D environments) raised at a ~$1.45B valuation in a round led by Natural Capital, with Amazon, AMD Ventures, GV, EQT, and In-Q-Tel joining. AWS becomes preferred cloud and will supply Trainium chips — a notable AWS/AMD win in a space NVIDIA had funded. (Irish Times | TFN)

  • ChatGPT's AI-assistant market share drops below 50% (Jun 16) Trending — By one widely cited measure ChatGPT fell to ~46% (from 87% a year ago), with Gemini at ~28% and Claude at ~10%; web-visit and MAU measures still put ChatGPT ahead (~55% / 1.1B users). Methodologies differ, but the trend — a tightening multi-model market — is consistent. Worth factoring into model-portfolio and procurement decisions. (TechCrunch)

AI in Production

  • GitHub's AI-agent reliability crisis pushes Microsoft to lean on AWS (Jun 16) Production — Update to the recurring Copilot-reliability thread. Following GitHub's nine May incidents and sub-99% June availability, reporting says Microsoft is tapping AWS capacity to stabilize AI-agent workloads as enterprise SLAs break. Reinforces this month's lesson: treat embedded AI assistants as infrastructure dependencies and plan for graceful degradation. (TechTimes)

Policy & Safety

  • AI dominates the G7 summit in France (Jun 15–17) Policy — At the Évian-les-Bains summit, OpenAI's Altman, Anthropic's Amodei, and DeepMind's Hassabis attended together for the first time, joining a working session on safe, rapid AI deployment. Agenda spanned frontier-AI risk, infrastructure/sovereignty, US industry dominance, and child online safety — a sign the public-policy layer is catching up to the labs. (CNBC | Bloomberg)

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