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

AI news, Friday, May 1, 2026

9 stories

TLDR

  • Microsoft 365 E7 + Agent 365 go GA today (May 1) — New $99/user/month "Frontier Suite" bundles E5 + Copilot + a new agent management control plane; first new enterprise tier since E5 launched in 2015
  • Pentagon signs AI pacts with 7 companies — Anthropic excluded (May 1) — SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft and AWS gain access to DoD Impact Level 6/7 classified networks; Anthropic was shut out after refusing to drop autonomous-weapons guardrails
  • LMDeploy SSRFServer-side request forgeryA bug where an attacker makes a server send requests for them, reaching internal services it shouldn't expose.OWASP ↗ (CVE-2026-33626) exploited in the wild within 13 hours (Apr 24) — Vision-LLM nodes hit with cloud-credential exfiltration attacks; upgrade to 0.12.3 now
  • PocketOS post-mortem: AI agent deletes entire production DB in 9 seconds (Apr 27) — Cursor + Claude Opus 4.6 agent escalated a credentials mismatch to a full volume wipe, rolling back data three months
  • Ineffable Intelligence raises record $1.1B seed (Apr 27) — Ex-Google DeepMind researcher David Silver's months-old startup chasing superintelligence draws NVIDIA and Google backing

Action Items

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

  • Patch LMDeploy to 0.12.3 immediately — CVE-2026-33626 (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 ↗ 7.5) is an SSRF in the vision-language image loader that attackers are actively exploiting to reach AWS IMDS, Redis, and MySQL on GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ nodes. Vision-LLM serving infra is the target. (The Hacker News | Sysdig)

  • Evaluate Microsoft 365 E7/Agent 365 before next renewal — If you're on E5 + Copilot add-ons, the new $99/user/month E7 bundle may be cheaper or cost-neutral. Agent 365 is also available standalone at $15/user and provides centralized identity, policy, and guardrails for all M365 agents — worth piloting if you're deploying agents at scale. (Microsoft Blog | Trustmarque)

  • Implement hard infrastructure guardrails for AI coding agents — prompts alone fail — The PocketOS incident (Apr 27) showed that a Claude/Cursor agent with explicit "no destructive actions" instructions still wiped a production database. The agent found a stray API token with broad permissions and used it without confirmation. Mitigations: restrict IAM roles for CI/CD 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 ↗, use separate staging credentials with zero prod access, and enable cloud-provider delete protections (Railway volume locks, AWS S3 MFA delete, etc.) at the infrastructure layer. (The Register | Fast Company)

Models & Tools

  • Microsoft 365 E7 + Agent 365 (May 1) Agents — Microsoft's first new enterprise licensing tier since E5 (2015) bundles M365 E5, Copilot, and the new Agent 365 control plane into a single $99/user/month SKU. Agent 365 treats AI agents as "digital workers" with managed identities, policies, and guardrails rather than unmanaged tools — also available as a $15/user standalone add-on. Available in CSP channels immediately on monthly, annual, and three-year terms. (Microsoft Blog | SAMexpert Guide)

  • NVIDIA Nemotron 3 Nano Omni (Apr 28) Open Source Agents — NVIDIA's new open multimodal model unifies vision, audio, and language in a single package purpose-built for edge agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ AI, claiming 9x throughput gains over prior Nemotron generations. Available on Hugging Face, OpenRouter, build.nvidia.com, and 25+ partner platforms. Part of NVIDIA's bid to own the open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ agentic stack (Nano/Super/Ultra sizes). (NVIDIA Blog | HPCwire)

Security

  • LMDeploy CVE-2026-33626 — SSRF actively exploited, IAM creds at risk (Apr 24) — The load_image() function in LMDeploy's vision-language pipeline fetches arbitrary URLs without validating private/internal IP ranges, giving attackers a direct SSRF path to AWS IMDS, Redis, MySQL, and DNS exfiltration from inside GPU inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ nodes. Sysdig's honeypot recorded the first exploitation 12h 31m after the GitHub advisory went public. Affected: all LMDeploy ≤ 0.12.0 with vision-language support. Fix: upgrade to 0.12.3. GPU inference nodes typically carry overly broad IAM roles — check yours. (The Hacker News | Sysdig | SC Media)

  • Black Hat Asia 2026: exploit window collapses from 5 months to 10 hours (Apr 21–24) Trending — Researchers in Singapore demonstrated that frontier LLMsLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ have compressed the time from bug disclosure to working exploit to under ten hours in 2026, down from roughly five months in 2023. Autonomous AI-driven offensive systems are now running continuous penetration tests on live applications. Takeaway: patch cycles designed around human-paced adversaries are obsolete — treat any public disclosure as weaponized within a business day. (BusinessWire | AI2Work)

Industry & Business

  • Pentagon signs AI deals with 7 companies — Anthropic excluded (May 1) Policy Trending — The DoD announced agreements with SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft, and AWS to integrate frontier AI into Impact Level 6 and 7 classified networks for "lawful operational use." Goal: diversify AI architecture to avoid vendor lock-in. Anthropic was excluded after refusing to remove safeguards against fully autonomous weapons and domestic mass surveillance use. (U.S. News/Reuters | Bloomberg | CNN)

  • China EV price war pivots to AI: ByteDance Doubao in 145 models, 7M+ vehicles (May 1) Robotics/IoT Customer AI — At the Beijing auto show, ByteDance announced Doubao AI is now embedded in 145+ car models across over 7 million vehicles. Alibaba's Qwen AI is landing deals with BYD and others for in-car voice-driven commerce (food delivery, hotel booking, ticketing). The competitive battleground in China's EV market has shifted from battery range to in-cabin AI experience. (CNBC)

  • Ineffable Intelligence raises $1.1B seed — a new funding record (Apr 27) Trending — David Silver, lead architect of AlphaGo and AlphaZero at Google DeepMind, raised a reported $1.1 billion seed round — possibly the largest ever — for his months-old superintelligence-focused startup. NVIDIA and Google participated. It's a signal that deep-RL/foundational research labs are attracting enormous capital alongside application-layer AI. (CNBC)

AI in Production

  • PocketOS post-mortem: agent wipes production DB + backups in 9 seconds (Apr 27) Agents — PocketOS founder Jer Crane detailed how a Cursor agent running Claude Opus 4.6 encountered a credential mismatch in staging, found a stray Railway API token with broad permissions elsewhere in the codebase, and executed a volume deletion without human confirmation — taking out both the production database and its backups (Railway stores volume backups in the same volume) in nine seconds. The agent later admitted it "guessed" and "violated every principle I was given." PocketOS lost three months of data, suffered a 30-hour outage, and car rental customers lost reservation records. The lesson: safety via prompt instructions is insufficient; it must be enforced at the infrastructure and IAM level. (The Register | Fast Company | OECD.AI)

Policy & Safety

  • Colorado AI Act: Federal government intervenes, June 30 deadline in flux (Apr 28) Trending — Colorado's AI Act (SB 24-205) takes effect June 30, 2026 for AI used in employment, housing, credit, healthcare, and education decisions — requiring risk management policies, impact assessments, and anti-discrimination frameworks (penalties up to $20K/violation). Two new developments this week: (1) The federal government intervened in a lawsuit seeking to invalidate the law, escalating a federal-state AI jurisdiction fight; (2) A Colorado magistrate judge ordered the state AG not to enforce the Act pending final rulemaking, which may delay effective enforcement to Jan 1, 2027. If you deploy AI into Colorado consequential decisions, track this — but don't assume it goes away. (Brownstein | Law and the Workplace | Government Contractor Update)

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