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

AI news, Sunday, May 3, 2026

11 stories

TLDR

  • Mistral Medium 3.5 drops (May 1) — Single dense 128B model unifies chat, reasoning, and code; $1.50/$7.50 per M 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 ↗. Vibe now ships async remote agents for GitHub/Slack.
  • Microsoft Agent 365 hits GA (May 1) — Enterprise agent governance layer live at $15/user/mo; new M365 E7 "Frontier Suite" ($99/user/mo) wraps it with Copilot + Entra.
  • Grok 4.3 launches at 40-60% price cut (Apr 30) — $1.25/$2.50 per M tokens, 1M context, native video input. Cheapest frontier-tier model with reasoning.
  • Cursor Security Review enters beta (Apr 30) — Always-on agents review every PR for vulns and run scheduled codebase scans; available on Teams/Enterprise now.
  • Meta raises 2026 AI capex to $125-145B; stock -7% (Apr 29) — Investors rattled as Zuckerberg couldn't answer the ROI question, despite a 33% revenue beat.

Action Items

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

  • Enable Cursor Security Review (Teams/Enterprise) — Available now as of Apr 30. Two always-on agents: Security Reviewer leaves inline comments on every PR; Vulnerability Scanner posts scheduled codebase findings to Slack. Both accept custom instructions and MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ server plugins. Configure via Cursor dashboard.
  • Benchmark Mistral Medium 3.5 for your coding/reasoning workloads — At $1.50 in / $7.50 out per M tokens, it's competitive with Claude Sonnet-tier pricing while unifying chat + reasoning + code into one model call with toggleable reasoning. May reduce model-switching overhead in agent pipelines.
  • Colorado AI Act compliance audit — June 30 deadline — The DOJ intervened Apr 28 in a lawsuit that could block or delay the Act, and Colorado is drafting a replacement framework (possible Jan 2027 date). Prepare compliance work now, but watch for injunction news before spending deeply.

Models & Tools

  • Mistral Medium 3.5 (May 1) Open Source — 128B dense model with toggleable reasoning and 256K-token context. Replaces separate chat/code/reasoning modelsA model that works through hidden reasoning steps before its final answer. Those steps are billed as output.OpenAI ↗ with a single call. API: $1.50 in / $7.50 out per M tokens. Vibe (Mistral's dev IDE) now ships async remote agents that run in isolated sandboxes with GitHub and Slack connectors — approval-required before sensitive actions. (The Decoder | Mistral Blog)

  • Grok 4.3 (Apr 30) Pricing — xAI cut input ~40% and output ~60% vs Grok 4.20, landing at $1.25/$2.50 per M tokens. Adds 1M-token context and native video input. Also ships a new fast voice-cloning suite. The 1M context + reasoning combination now costs roughly 1/12th of Claude Opus 4.7. (VentureBeat | xAI Docs)

  • Microsoft Agent 365 GA (May 1) Agents — Enterprise control plane for managing all AI agents across M365. Now governs both cloud agents and locally running agents on Windows endpoints. Priced at $15/user/mo standalone, or bundled in the new M365 E7 Frontier Suite at $99/user/mo (includes Copilot + Entra Suite). (Microsoft Security Blog | Trustmarque)

  • Cursor Security Review beta (Apr 30) Security Agents — Available on Teams and Enterprise plans. Security Reviewer checks every PR for vulns, auth regressions, privacy/data-handling risks, and prompt injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ — leaves inline diff comments with severity and remediation. Vulnerability Scanner runs scheduled scans of your codebase and posts to Slack. (Cursor Changelog)

APIs & Pricing

  • Grok 4.3: 40-60% API price cut (Apr 30) — Input: $1.25/M (was ~$2.00+), Output: $2.50/M (was ~$6.25). Undercuts comparable reasoning-capable models across the board. See Models & Tools above for context. (xAI Docs)

  • DeepSeek V4 Pro promo floor expires May 5 — In 2 days, the $0.435/$0.87 launch rate briefly disappears before a 75%-off promo applies through May 31. If you haven't run a quick benchmark, today is the last full weekend. (Covered in full in yesterday's digest.)

Security

  • Three AI coding agents leak secrets via single prompt injection (~Apr 2026) Agents — Researchers showed Cursor, the Claude Code GitHub Action, and GitHub Copilot can all be manipulated via crafted prompt injection to exfiltrate secrets to attacker-controlled endpoints. Anthropic classified the Claude Code finding as 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 ↗ 9.4 Critical — and Anthropic's own system card disclosed the GitHub Action "is not hardened against prompt injection." Google paid $1,337; GitHub paid $500. Separately: a VentureBeat survey found 65.3% of organizations have deployed zero dedicated prompt injection defenses. (VentureBeat)

Industry & Business

  • Meta Q1 2026: Revenue beats, AI capex balloons to $125-145B (Apr 29) Infra — Meta posted $56.3B revenue (+33% YoY) and $26.8B net income, beating all estimates. But it raised 2026 capex from $115-135B to $125-145B for AI infrastructure; stock fell ~7%. Zuckerberg, asked for ROI signals, called it "a very technical question." Microsoft and Google also raised AI capex this week — only Google convinced investors its spend is paying off. Combined, the five hyperscalers are on track to spend $700B on AI infrastructure in 2026. (CNBC | Fortune | Fortune)

AI in Production

  • GitHub discloses outages tied to AI-driven scale and architectural coupling (Apr 2026) Infra — GitHub publicly acknowledged it failed its own reliability standards in recent months, attributing repeated disruptions to rapid growth, tight service coupling, and insufficient load-handling headroom. A useful post-mortem signal for AI-adjacent platforms: prioritize bulkhead isolation and load-shedding before traffic surprises you. (InfoQ)

Policy & Safety

  • DOJ intervenes to potentially block Colorado AI Act (Apr 28) — The federal government filed to intervene in xAI's lawsuit challenging Colorado's AI Act. If successful, it could halt or delay the June 30, 2026 compliance deadline. Colorado is simultaneously drafting a replacement framework that would push the effective date to Jan 1, 2027. The June 30 deadline still stands — but a stay is now plausible. (Government Contractor Compliance Update | Law and the Workplace)

  • China finalizes anthropomorphic AI interaction rules (Apr 10, effective Jul 15) Policy — Interim Measures for the Management of Anthropomorphic AI Interaction Services take effect July 15, 2026. Requires companion/emotional AI apps to: disclose AI nature at all times, implement addiction monitoring and emotion-state checks, obtain guardian consent for minors, and take crisis-intervention steps for at-risk users. Bans virtual intimate relationships with minors entirely. If you ship AI companion features into China, a 73-day compliance clock is running. (Global Times | Geopolitechs)

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