JadidJust-in AI Developments, Incidents & Deadlines
April 2026
SMTWTFS1234567891011121321415161718192021222324252627282930

AI news, Tuesday, April 14, 2026

13 stories

TLDR

  • Claude Mythos Preview (Apr 7) — Anthropic's most powerful model ever, 93.9% on SWE-benchA benchmark that tests whether a model can resolve real GitHub issues by editing code in popular Python repositories.Jimenez et al. 2023 ↗ Verified, but deliberately withheld from public release due to cybersecurity risks Trending
  • OpenAI acquires Hiro Finance (Apr 13) — acquihire of AI personal-CFO startup; Hiro shuts down April 20, data deleted by May 13 Business
  • GLM-5.1 tops SWE-Bench Pro (Apr 7) — Z.ai's open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ 744B 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 ↗ beats Claude Opus 4.6 and GPT-5.4 on the hardest agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ coding benchmark; MIT license Open Source
  • MiniMax MMX-CLI + Music 2.6 (Apr 10–12) — multi-modal agent CLI and open-sourced MusicSkills hit this week Agents Open Source
  • 5 days left: Claude 3 Haiku retires April 19 — final reminder to migrate to claude-haiku-4-5

Action Items

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

  • Migrate off claude-3-haiku-20240307 NOW — 5 days left before April 19 retirement. Drop-in replacement is claude-haiku-4-5. (Anthropic Deprecations Docs)
  • Evaluate GLM-5.1 for agentic coding workloads — Z.ai's open-weight 744B MoE is now #1 on SWE-Bench Pro (58.4), edging Claude Opus 4.6 and GPT-5.4. MIT license; weights on Hugging Face. Worth benchmarking on your own tasks before assuming a proprietary model is the only option. (Z.ai Docs | GitHub)
  • Patch ActiveMQ if used in MLOps pipeline — CVE-2026-34197 is an RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ via the Jolokia API in Apache ActiveMQ. If your data pipelines or agent orchestration uses ActiveMQ for messaging, treat this as critical. (Horizon3.ai)

Models & Tools

  • Claude Mythos Preview (Apr 7) Trending — Anthropic's most capable model achieves 93.9% on SWE-bench Verified, 77.8% on SWE-bench Pro, and autonomously discovered thousands of zero-daysAn attack using a flaw that was unknown before it was exploited.NIST ↗ across major OSes and browsers. Anthropic deliberately will not release it publicly; access restricted to "Project Glasswing" — a 12-company consortium (Amazon, Apple, Google, Microsoft, NVIDIA, and others) using it exclusively for defensive cybersecurity, backed by $100M in Anthropic usage credits. (InfoQ | NxCode)

  • GLM-5.1 (Z.ai) (Apr 7) Open Source — Z.ai (formerly Zhipu AI) releases a 744B Mixture-of-Experts model with 40B active parameters, 200K context, and MIT license. Scored 58.4 on SWE-Bench Pro — the first open-weight model to top that benchmark, beating Claude Opus 4.6 (57.3) and GPT-5.4. Weights on Hugging Face. (Dataconomy | GitHub)

  • MiniMax MMX-CLI (Apr 12) Agents Open Source — Node.js CLI that wraps MiniMax's full generative suite (text, image, video, speech, music, vision, search) into standard shell commands. AI agents get native multi-modal access with zero MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ configuration. (MarkTechPost)

  • MiniMax Music 2.6 + MusicSkills (Apr 10) Agents Open Source — Music 2.6 adds a Cover feature and reduces generation latency. Three MusicSkills open-sourced for the agent ecosystem: minimax-music-gen (prompt-to-track), minimax-music-playlist (personalized curation), and buddy-sings (persona-based singing). Compatible with agent frameworks via MMX-CLI. (KuCoin)

  • GigaTIME in Azure AI Foundry Labs (Apr 2026) Open Source Infra — Microsoft's open-source cancer-mapping AI (co-developed with Providence, originally open-sourced Dec 2025) is now integrated into Azure AI Foundry Labs. Translates standard $5–10 H&E pathology slides into virtual multiplex immunofluorescence images across 21 protein channels, trained on 40M cells from 14K+ patients. Newly accessible to researchers via Foundry Labs and Hugging Face. (Microsoft Foundry Labs | Microsoft Blog)

APIs & Pricing

  • Claude 3 Haiku — April 19 (5 days) — claude-3-haiku-20240307 stops working in 5 days. This is the final digest reminder before the cutoff. Migrate to claude-haiku-4-5. (Anthropic)

Security

  • CVE-2026-34197: Apache ActiveMQ RCE via Jolokia API (Apr ~10) — Critical remote code execution vulnerability in the Jolokia management endpoint of Apache ActiveMQ. MLOps pipelines that use ActiveMQ for message brokering between services are directly exposed. Patch immediately. (Horizon3.ai)

Industry & Business

  • OpenAI acquires Hiro Finance (Apr 13) — OpenAI acquihires the AI-powered personal finance startup (backed by Ribbit, General Catalyst). Hiro team joins to build ChatGPT financial planning features. Hiro stops accepting signups today (Apr 14), product shuts down Apr 20, all user data deleted by May 13. Terms undisclosed. (TechCrunch | The Decoder)

  • Q1 2026 VC: $300B record, 80% AI (Apr) — Global venture funding hit $300B in Q1 — up 150%+ YoY — with AI absorbing $242B (80% of total). Four of the five largest VC rounds ever recorded closed in Q1: OpenAI ($122B), Anthropic ($30B), xAI ($20B), Waymo ($16B). Defense AI (Shield AI $1.5B Series G at $12.7B valuation) and AI compute infra round out the top categories. (Crunchbase | Crunchbase - Foundation AI)

  • NVIDIA Quantum Day (Apr 14, today) Infra — Free virtual event running 8 AM–12:45 PM PT. Focus: AI-accelerated quantum error correction, hybrid quantum-GPU workflows, and DGX Spark for quantum labs. Worth a watch for teams thinking about quantum-classical integration. (NVIDIA)

AI in Production

  • Bluesky AppView cascade failure post-mortem (Apr 4–7) — Root cause: memcached errors triggered unbounded error logging at millions of req/sec → Go runtime spawned 10× OS threads (150 → 1,500) → massive GC stop-the-world pauses → full service outage. Lesson: logging on error paths in high-throughput hot loops is a silent latency bomb. Instrumentation must be carve-out, not on the critical path. (Hacker News thread)

Policy & Safety

  • California Governor EO on AI procurement (Apr 2026) — Executive Order sets AI procurement standards and responsible-use requirements for California state government agencies. Practical implications for enterprise AI vendors selling to state government; review if your products are in that channel. (DLA Piper)

  • EU AI Act binding provisions: August 2026 — Most enforcement-carrying rules of the EU AI Act are expected to come into force in August 2026, covering high-risk AI applications in healthcare, HR, and finance. Documentation, transparency, and mandatory human oversight requirements apply to deployers — not just developers. Start gap assessments now if your product touches EU users in these verticals. (OneTrust | Wilson Sonsini)

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