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

AI news, Saturday, May 9, 2026

9 stories

TLDR

  • Semantic Kernel RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ + file-write CVEsCommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗ disclosed (May 7) — Prompt injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ in SK Python (<1.39.4) enables host RCE via eval(); SK .NET (<1.71.0) allows arbitrary file writes. Upgrade now.
  • CAISI signs pre-release evaluation deals with Google DeepMind, Microsoft, xAI (May 5) — Trump administration formalizes frontier model vetting; WH drafting executive order to make it mandatory for all labs.
  • Moonshot AI closes $2B at $20B valuation (May 7) — Kimi K2.6 (1T-param open-weightsA model whose trained weights are published for anyone to download and run.NTIA ↗) now #2 on OpenRouter; Moonshot up 5× in valuation over six months.
  • Gartner: 80% of companies cut staff via AI — most saw negative ROI (May 6) — Organizations that amplify human capacity outperform those that replace workers.
  • NVIDIA Nemotron 3 Nano Omni (May 5) — Open multimodal model (vision + audio + language) promising up to 9× throughput gains for agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ workloads.

Action Items

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

  • Upgrade Semantic Kernel now (CVE-2026-25592 + CVE-2026-26030) — SK Python to ≥1.39.4, SK .NET to ≥1.71.0. CVE-2026-26030 lets a single crafted prompt achieve host RCE via an eval()-based vector filter; CVE-2026-25592 exposes an unguarded file-download helper that allows arbitrary host file writes. Both are exploitable through prompt injection with no additional auth required. Disclosed May 7. (Microsoft Security Blog | NVD-26030 | NVD-25592)
  • DALL-E 3 API shuts off in 3 days (May 12) — final window — The gpt-image-1 migration is not parameter-compatible; if you haven't tested it yet, do it today or accept broken image pipelines on Monday.
  • Colorado AI Act effective June 30 — Fifty-three days until compliance is required for AI used in consequential decisions (hiring, credit, housing, etc.). The law mandates risk management policies, impact assessments, and anti-discrimination notices. Loop in legal now if you haven't. (Cooley overview)

Models & Tools

  • NVIDIA Nemotron 3 Nano Omni (May 5) [Open Source] [Infra] — NVIDIA released a new open multimodal model that unifies vision, audio, and language in a single architecture, designed specifically for on-device and edge-inference agent scenarios. NVIDIA claims up to 9× higher throughput vs. Nemotron 2 Nano. The broader Nemotron 3 family (Nano, Super, Ultra) is being released in stages through H1 2026. Nano Omni is the first multimodal variant to ship. (NVIDIA Blog | NVIDIA Newsroom)

  • UiPath Automation Suite: on-prem agentic AI for regulated industries (May 5) [Agents] [Infra] — UiPath shipped agentic AI capabilities for its self-hosted Automation Suite, specifically targeting government agencies and industries with data sovereignty constraints. Supports self-hosted or cloud LLMsLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ (OpenAI, Gemini, Anthropic), runs on AWS, Azure, and OpenShift, and arrives with ISO/IEC 42001 and FedRAMP compliance built in. For teams evaluating agentic deployments in air-gapped or restricted cloud environments, this is the most production-ready option in its category. (BusinessWire | Help Net Security)

Security

  • Microsoft Semantic Kernel: prompt injection → RCE and file-write (CVE-2026-26030 + CVE-2026-25592) (May 7) [Agents] — Microsoft disclosed two patched vulnerabilities in Semantic Kernel — one of the most widely used open-source agent frameworks. CVE-2026-26030 (Python SDK < 1.39.4): the InMemoryVectorStore filter was implemented as a Python lambda executed via eval(), with user-controlled data interpolated directly into the expression; a single crafted prompt injection achieves host RCE. CVE-2026-25592 (.NET SDK < 1.71.0): DownloadFileAsync was accidentally exposed to the model as a callable kernel function without path validation, enabling a hostile prompt to write files anywhere on the host. Both flaws require a prompt injection entry point but no additional auth. Fix: SK Python ≥1.39.4 / SK .NET ≥1.71.0. The headline takeaway: enterprise AI risk has moved from the conversation window to the process tree. (Microsoft Security Blog | NVD CVE-2026-26030 | NVD CVE-2026-25592 | Windows Forum analysis)

Industry & Business

  • Moonshot AI closes $2B at $20B valuation (May 7) [Open Source] [Trending] — China-based Moonshot AI (makers of Kimi) closed a $2B round led by Meituan's Long-Z Investments, with Tsinghua Capital and China Mobile also participating. Valuation surged 5× in six months ($4.3B → $20B). Its latest open-weights model, Kimi K2.6 (~1T parameters, 384 MoEMixture of expertsA model made of many sub-networks, called experts, where a learned router sends each input to only a few of them.Shazeer et al. 2017 ↗ experts), is now the #2 most-used LLM on OpenRouter behind GPT-5.5. Moonshot's ARR crossed $200M in April, making it the most credible open-source challenger to DeepSeek in the coding/agent domain. (TechCrunch | SiliconAngle)

  • Freshworks cuts 11%; Coinbase cuts 14% (May 5) [Hiring] — Both companies cited AI-driven productivity shifts: Freshworks (~500 employees) with its CEO publicly stating AI now writes "over half" of the company's code; Coinbase (~700 employees) shifting toward an agent-first workflow. These two join a 2026 YTD total exceeding 95,000 tech layoffs, running at 864/day — faster than 2025's pace. (InformationWeek tracker)

  • Gartner: AI layoffs backfiring at 80% of companies (May 6) [Hiring] [Trending] — A survey of 350 global enterprises (>$1B revenue) found 80% had reduced headcount as a direct result of AI deployment — but most reported negative ROI. Gartner's VP Analyst: "Workforce reductions may create budget room, but they do not create return." A parallel warning from Imperial College London and Microsoft: AI adoption is paradoxically increasing worker burden as employees spend time supervising and correcting AI agents. Organizations actually seeing positive returns are those investing in new roles and human-AI teaming models. (The Register)

  • Anthropic $50B round nearing close (ongoing) [Trending] — Anthropic is in late-stage discussions on a $50B fundraise at a ~$900B valuation (ARR reported at $30B+); round expected to close within the next two weeks per earlier reporting. Dragoneer, General Catalyst, and Lightspeed are among interested investors; not yet confirmed. At $900B, Anthropic would surpass OpenAI's $852B post-money valuation from its March round. (TechCrunch | Bloomberg)

Policy & Safety

  • CAISI formalizes pre-release AI evaluation with Google DeepMind, Microsoft, xAI (May 5) [Trending] — The Center for AI Standards and Innovation — formerly AISI (the word "safety" was intentionally removed) — signed pre-deployment evaluation agreements with Google DeepMind, Microsoft, and xAI, extending existing deals with OpenAI and Anthropic. CAISI will conduct capability assessments and security research before public model releases. The Trump White House is separately drafting an executive order that would make pre-release model vetting mandatory across all frontier labs — a sharp reversal from the administration that once dismantled Biden's AI oversight apparatus. The catalyst: Anthropic's Claude Mythos autonomously discovered thousands of zero-dayAn attack using a flaw that was unknown before it was exploited.NIST ↗ vulnerabilities in critical infrastructure during internal testing, alarming White House officials. (CNBC | Fortune | Axios)

  • Colorado AI Act: 53 days to compliance (June 30, 2026) (Policy) — One of the most comprehensive state AI laws currently in effect in the U.S. applies to AI used as a "substantial factor" in consequential decisions (employment, credit, housing, education, healthcare). Requires developers and deployers to conduct impact assessments, maintain AI risk management policies, implement anti-bias disclosures, and enable human review of automated decisions. (Cooley | Gunderson Dettmer analysis)

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