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

AI news, Friday, April 10, 2026

22 stories

TLDR

  • GLM-5.1 tops SWE-BenchA benchmark that tests whether a model can resolve real GitHub issues by editing code in popular Python repositories.Jimenez et al. 2023 ↗ Pro (Apr 7): Z.ai's open-source 754B 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 ↗ model scores 58.4 on SWE-Bench Pro, beating GPT-5.4 and Claude Opus 4.6, with 8-hour autonomous agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ task runs under MIT license
  • Anthropic three-day outage (Apr 6–8): Cascading inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ infrastructure failure disrupted enterprise workflows globally — post-mortem cites a load-balancing config gone wrong
  • PraisonAI triple CVECommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗ (Apr 8): Three RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ vulnerabilities (including a critical 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.x sandbox escapeBreaking out of a sandbox, the restricted environment that limits what untrusted code can reach.NIST ↗) affect versions prior to 4.5.115 — patch immediately if you use PraisonAI
  • Gemini API billing crackdown (Apr 1): Free tier gutted 50–80%, Pro models paywalled, spending caps enforced; Gemini 2.0 Flash deprecated June 1
  • Q1 2026 funding record (Apr): $297B in AI startup funding — more than 3× the prior quarter; tech layoffs hit 80K with ~50% attributed to AI automation

Action Items

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

  • Patch PraisonAI to ≥4.5.115 now — CVE-2026-39891 (CVSS 8.8, template injection RCE), CVE-2026-39890 (critical YAML parsing RCE), and CVE-2026-39888 (critical sandbox escape RCE) are all actively documented
  • Migrate off Gemini 2.0 Flash before June 1 — it's being deprecated; evaluate Gemini 2.5 Flash or Gemini 3.1 Flash-Lite as drop-in replacements; also audit free-tier Gemini Pro usage under the new April spending caps
  • Migrate off Claude 3 Haiku — deprecated April 2026; move workloads to Haiku 4.5 before cutoff
  • Add a fallback inference provider — Claude's 3-day outage (Apr 6–8) exposed a single-point-of-failure risk for teams relying solely on Anthropic; review your resilience posture

Models & Tools

  • GLM-5.1 (Apr 7) Open Source Trending — Z.ai (Zhipu AI) released this 754B MoE model under MIT license; it hit #1 on SWE-Bench Pro (58.4, edging past GPT-5.4's 57.7 and Claude Opus 4.6's 57.3) and is engineered for up to 8-hour autonomous agentic task loops. Weights are on HuggingFace at zai-org/GLM-5.1; full writeup at VentureBeat

  • Google Gemma 4 (Apr 2) Open Source — Four-variant family (E2B, E4B, 26B MoE, 31B Dense) released under Apache 2.0; handles text, images, and audio natively with 256K context and runs locally on edge hardware. Competitive with models far above its parameter count. (Google DeepMind | Google Blog)

  • MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ crosses 97M monthly installs (Mar stat, trending now) Trending — Anthropic's Model Context Protocol hit 97M SDK downloads/month in March 2026 (up from ~2M at Nov 2024 launch) and now has 5,800+ community servers; every major AI provider (OpenAI, Google, Microsoft, AWS) ships MCP-compatible tooling. The protocol war is effectively over. (The New Stack | DEV Community summary)

  • Microsoft 2026 Release Wave 1 (Apr) Agents — Expanded agentic AI across Dynamics 365, Power Platform, and M365 Copilot with deeper autonomous workflow capabilities and governance controls. Worth reviewing if your org runs any M365 Copilot deployments. (Cloud Wars)

  • Google Colab MCP Server (Apr) Agents — Google launched an official MCP server for Colab, enabling any MCP-compatible AI agent to connect to Colab notebooks and run Python directly. Low-friction path for adding code execution to agent pipelines. (Google Developers Blog)

APIs & Pricing

  • Gemini 2.0 Flash / Flash-Lite: Deprecated June 1, 2026 (Apr) — Migration deadline is firm; move to Gemini 2.5 Flash or Gemini 3.1 Flash-Lite. Free-tier quotas for Gemini API were already cut 50–80% in December; April 1 enforced mandatory spending caps ($250/mo Tier 1, $2K Tier 2) and removed Pro model access for free users entirely. (FindSkill.ai guide | Detailed billing changes)

  • Claude 3 Haiku: Deprecated April 2026 (Apr) — Legacy Haiku is retiring this month. Migrate workloads to Haiku 4.5 to avoid disruption. Anthropic also quietly dropped Opus pricing 67% and extended context to 1M 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 ↗ — check if you're still on old pricing tiers. (LLM API Pricing overview)

  • OpenAI Videos API / Sora 2: Deprecating Sept 24, 2026 (Mar 24) — If you use Sora 2 model aliases or the Videos API, you have until September 24 to migrate. (OpenAI Deprecations)

Security

  • PraisonAI: Three RCE CVEs (CVE-2026-39891, 39890, 39888) (Apr 8) Trending — All three affect PraisonAI < 4.5.115. CVE-2026-39891 (CVSS 8.8): template injection via unescaped agent input → arbitrary code execution. CVE-2026-39890 (critical): RCE via malicious YAML parsing. CVE-2026-39888 (critical): sandbox escape RCE. If PraisonAI is anywhere in your stack, upgrade to 4.5.115 now. (TheHackerWire CVE-2026-39891 | GitLab Advisory)

  • Prompt injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ entering government/enterprise daily workflows (Apr 9) — Help Net Security reports that as GenAI moves into daily operations, indirect prompt injection (malicious instructions embedded in emails, web pages, or docs that AI later retrieves) is now OWASP's #1 GenAI risk category. Worth auditing any RAGRetrieval-augmented generationFetching relevant documents at query time and giving them to the model, so its answer draws on them.Lewis et al. 2020 ↗ pipelines that ingest untrusted external content. (Help Net Security)

Industry & Business

  • Oracle: ~30,000 layoffs to fund AI infrastructure (Mar 31) Hiring — Oracle cut roughly 18.5% of its global workforce in a single email blast at 6am on March 31, with ~12,000 cuts in India. The company disclosed a $2.1B restructuring plan in its Q3 10-Q; it's trading human headcount for data center CapEx, not cutting due to financial distress (net income up 27% YoY). (CNBC)

  • Q1 2026 tech layoffs: 80K jobs, ~50% AI-attributed (Apr) Hiring — Around 78,557 tech workers were cut in Q1 2026, with ~37,638 positions (47.9%) explicitly attributed to AI automation. The "cut and redirect" pattern (e.g., Atlassian cutting 1,600 roles while hiring 800 AI-focused roles) is the defining trend. AI/ML engineering roles are seeing 163% YoY growth in job postings. (TechRadar | Tom's Hardware)

  • Eclipse VC raises $1.3B for robotics and AI infrastructure (Apr 7) — Cerebras backer Eclipse closed a $1.3B fund targeting physical AI, manufacturing, and defense infrastructure — a signal of where serious capital is moving beyond foundation models. (Bloomberg)

  • Q1 2026: Record $297B AI startup funding (Apr) — North American AI startups raised a record $297B in Q1 2026, more than 3× the prior quarter; foundational AI startup funding alone was double all of 2025. 87%+ of Q1 venture went to AI-related categories. (Crunchbase)

  • SpaceX-xAI merger in IPO spotlight (Apr 7, trending) Trending — The February 2026 SpaceX acquisition of xAI (valuing xAI at $250B) is back in the news: Musk is reportedly requiring banks working on SpaceX's planned $1.5–1.75T IPO to subscribe to Grok. A governance concern worth watching if your org does business with these banks. (PYMNTS)

AI in Production

  • Anthropic Claude: Three-day cascading outage (Apr 6–8) Infra — Claude suffered a 10-hour outage on April 6 — a routine load-balancing config triggered a feedback loop that overwhelmed GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ request queues. Similar elevated errors continued April 7 and 8, peaking at 3,000+ Downdetector reports. The post-mortem was technically detailed but vague on remediation commitments. Key lesson: any enterprise critical-path workflow running solely on Claude needs a fallback provider. (WebProNews | AI Daily)

Policy & Safety

  • Colorado AI Act takes effect June 30, 2026 (Apr) — The Colorado AI Act will impose substantial new obligations on AI developers and deployers including algorithmic discrimination requirements and impact assessment duties. If your product touches Colorado users, your legal team should be reviewing compliance now. (Global Policy Watch Q1 recap)

  • State therapy chatbot bans advancing (Apr 9–10) — Maine's LD 2082 (banning clinical AI use in mental health therapy) is heading to the governor; Missouri has a similar provision in an omnibus health care bill. If your team builds or integrates AI mental health tooling, this regulatory direction is worth tracking. (Transparency Coalition)

  • EU AI Act high-risk system rules expected August 2026 (Apr) — The first binding enforcement deadlines for high-risk AI systems under the EU AI Act are now expected to land in August 2026. EU-market deployments in healthcare, employment, or critical infrastructure should be in compliance readiness mode. (Pearl Cohen law firm overview)

Benchmarks & Evals

  • GLM-5.1 leads SWE-Bench Pro; leaderboard reshuffled (Apr 7) — SWE-Bench Pro scores as of this week: GLM-5.1 (58.4, open source) > GPT-5.4 (57.7) > Claude Opus 4.6 (57.3). On SWE-bench Verified, multiple sources report Claude Sonnet 4.6 around 79–80% and Gemini 3.1 Pro around 78.8%. Score variation across leaderboards is significant — methodology matters. (SWE-bench.com | BenchLM.ai)
  • zai-org/GLM-5 — GLM-5.1 weights and code; #1 open source on SWE-Bench Pro this week, 754B MoE under MIT license
  • google-gemini/gemini-cli — Google's open-source terminal AI agent with MCP support, 1M context, and Apache 2.0 license; free with personal Google account

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