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

AI news, Monday, April 27, 2026

9 stories

TLDR

  • Microsoft & OpenAI go non-exclusive (Apr 27) — Microsoft drops exclusive model access; OpenAI can now serve all its products on any cloud. Revenue-sharing capped and restructured through 2030. MSFT shares fell ~3%.
  • China blocks Meta's $2B Manus acquisition (Apr 27) — China's NDRC ordered the deal unwound, citing AI technology leakage concerns. Manus was already integrated into Meta's systems.
  • MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ design flaw enables RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ across the AI agent ecosystem (Apr 20) — Architectural bug in all Anthropic MCP SDKs exposes ~200,000 servers; Cursor, Claude Code, VS Code, Windsurf all affected. Anthropic called it "expected behavior."
  • Anthropic publishes Claude Code quality regression post-mortem (Apr 23) — Three engineering changes between March 4 and April 16 caused the widely-reported performance drop. All three are now reverted.
  • Alibaba Happy Horse video model enters beta (Apr 27) — Top-ranked text-to-video model hits public beta; available via the Happy Horse website and Alibaba Cloud API.

Action Items

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

  • Audit all MCP server instances for the RCE design flaw — OX Security found that Anthropic's official MCP SDKs (Python, TypeScript, Java, Rust) allow arbitrary command execution by design: any MCP server can be weaponized to gain access to API keys, databases, chat histories, and internal systems. Anthropic declined to patch the protocol architecture. Affected tools include Claude Code, Cursor, VS Code (Copilot), Windsurf, and Gemini-CLI. If you run MCP servers in CI, staging, or production, assess your exposure now. (The Hacker News | OX Security)

  • Migrate off Claude 3 Haiku before end of April — Legacy Claude 3 Haiku ($0.25/$1.25 per MTok) is being retired this month. Migrate workloads to Haiku 4.5 before the cutoff to avoid breakage. Check your API usage dashboards now.

  • Reassess OpenAI vendor lock-in strategy — With the Microsoft-OpenAI exclusivity gone, OpenAI can now deliver its models directly to AWS, Google Cloud, and other providers. If your org has been betting on Azure as the single OpenAI access point, this opens up multi-cloud options — but also means the OpenAI ecosystem is fragmenting. Evaluate whether direct OpenAI API access or an alternate cloud route is now advantageous. (Microsoft Blog | CNBC)

Models & Tools

  • Alibaba HappyHorse-1.0 (Apr 27) Open Source — After topping the Artificial Analysis blind-test leaderboard for text-to-video and image-to-video (1389 Elo, ~115 points above ByteDance's Seedance 2.0), Alibaba's ATH unit opened public beta. Available via the Happy Horse website and Alibaba Cloud Model Studio API. Offers text-to-video and image-to-video in a unified pipeline with joint audio-video generation and 1080p clips at 5–8 seconds. (CNBC | Bloomberg | GURUfocus)

APIs & Pricing

  • Claude 3 Haiku EOL — April 2026 Infra — Legacy Claude 3 Haiku is being deprecated this month. Migrate to Haiku 4.5, which offers better performance at competitive pricing. Check Anthropic's status page and your API dashboard for your specific cutoff date.

  • Gemini Pro models removed from free tier (Apr 1) Infra — Reminder: Gemini 3.1 Pro, 3 Pro, and 2.5 Pro are now paid-only on the Gemini API. Flash models remain free with reduced daily quotas. Gemini 2.0 Flash and 2.0 Flash-Lite will be fully deprecated June 1, 2026 — migrate to 2.5 Flash or 3 Flash. (FindSkill.ai | AI Pricing Guru)

Security

  • MCP "by design" RCE — all official Anthropic SDKs (Apr 20) Agents Infra — OX Security disclosed an architectural flaw baked into Anthropic's MCP reference SDKs (Python, TypeScript, Java, Rust): MCP servers can be weaponized to execute arbitrary commands on host systems, exposing API keys, databases, and chat histories. The flaw affects an estimated 150M+ downloads and 200,000 vulnerable instances across tools including Claude Code, Cursor, VS Code (with Copilot), Windsurf, and Gemini-CLI. Anthropic declined to modify the protocol architecture, calling the behavior "expected" — meaning developers who inherit the reference implementation must mitigate the risk themselves. (The Hacker News | OX Security | CSA Research)

Industry & Business

  • Microsoft-OpenAI partnership restructured — exclusivity ends (Apr 27) Trending — Microsoft's license to OpenAI IP remains through 2032 but becomes non-exclusive. OpenAI can now distribute its models to any cloud provider, including AWS and Google Cloud. Microsoft stops paying a revenue share to OpenAI; OpenAI continues paying ~20% to Microsoft through 2030, now capped. OpenAI products still launch first on Azure unless Microsoft declines. Market reaction: MSFT fell ~3%; GOOG and AMZN edged up. Practical implication: OpenAI will compete more aggressively on GCP and AWS, giving teams more flexibility on where to run OpenAI workloads. (Microsoft Blog | CNBC | Bloomberg)

  • China blocks Meta's $2B acquisition of Manus AI (Apr 27) Agents Policy — China's National Development and Reform Commission ordered Meta to unwind its acquisition of Manus, the Singapore-based general-purpose agent startup with Chinese founders, citing AI technology export concerns. Manus was already integrated into Meta's internal systems and its executives had joined Meta — making the unwind practically complex. First major case of China using export-control authority to block a US AI talent acquisition; signals Beijing is treating top-tier agent developers as national security assets. (CNBC | Bloomberg | NPR)

AI in Production

  • Anthropic post-mortem: three engineering changes caused Claude Code's month-long regression (Apr 23) Trending — Anthropic published a technical post-mortem identifying the three product-layer changes behind the widely-reported quality drop: (1) reasoning effort downgraded from high → medium on Mar 4 to reduce UI latency (reverted Apr 7); (2) a caching bug introduced Mar 26 caused Claude to forget context every turn instead of only after 1hr idle (fixed Apr 10); (3) overly strict verbosity limits added Apr 16 (≤25 words between tool calls, ≤100-word final responses) caused a 3% drop in coding evals (reverted Apr 20). Key operational lesson: small system-prompt changes can have large, hard-to-detect quality regressions in production AI tools — Anthropic is now requiring more internal staff to dogfood public builds to catch these faster. Anthropic also reset usage limits for all subscribers as remediation. (Anthropic Engineering | Fortune | VentureBeat)

Policy & Safety

  • China-Meta-Manus: first AI talent-acquisition block on national security grounds (Apr 27) — See Industry & Business above. Policy The NDRC's one-line statement blocking the deal is the clearest signal yet that Beijing intends to use export-control and investment-review authority to prevent Chinese-founded AI talent and IP from moving to US labs — regardless of whether the startup has technically relocated to a third country. Watch for similar actions against other Chinese-founded AI companies operating from Singapore or the UAE.

Benchmarks & Evals

  • 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 leaderboard — April 2026 snapshot — Claude Opus 4.7 leads at 87.6% (up from ~65% top score in early 2025), followed by GPT-5.3-Codex at 85.0% and Gemini 3.1 Pro at 80.6%. Critically: Claude Mythos Preview scores 93.9% on SWE-bench Verified but drops to 45.9% on SWE-bench Pro (the contamination-free version) — a 48-point gap that underscores why benchmark choice matters when evaluating agent coding capability for production use. (marc0.dev leaderboard | BenchLM.ai | CodeAnt analysis)

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