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

AI news, Wednesday, April 22, 2026

17 stories

TLDR

  • SpaceX gets an option to buy Cursor for $60B (Apr 21) — Partnership plus a right-to-acquire; if SpaceX walks, it pays $10B for work done. Supersedes yesterday's $2B / $50B fundraise story.
  • Anthropic + Amazon expand to 5GW compute + $25B new investment (Apr 20–21) — Anthropic commits $100B to AWS over 10 years; ~1GW of Trainium2/3 online by year-end. Claude Opus 4.7 now GA on Bedrock.
  • Jeff Bezos' Project Prometheus nears $10B at $38B valuation (Apr 21) — Physical-AI lab (5 months old), JPMorgan + BlackRock in; targets AI for aerospace/auto manufacturing.
  • OpenAI ships ChatGPT Images 2.0 (gpt-image-2) (Apr 21) — First OpenAI image model with native reasoning, 2K output, up to 8 consistent images per prompt; DALL-E 2/3 retire May 12.
  • Unauthorized group accessed Claude Mythos Preview (Apr 21) — Bloomberg/TechCrunch: a Discord group obtained access via a third-party vendor environment; Anthropic is investigating and says core systems are unaffected.

Action Items

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

  • Audit every MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ server and client in your stack against the "by-design" RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ issue — Researchers (OX Security, et al.) disclosed a systemic flaw in Anthropic's MCP SDKs (Python/TypeScript/Java/Rust) affecting 7K+ public servers and 150M+ downloads. Related CVEsCommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗ include CVE-2026-22252 (LibreChat), CVE-2026-22688 (WeKnora), CVE-2025-54136 (Cursor), CVE-2025-54994 (create-mcp-server-stdio), CVE-2025-49596 (MCP Inspector). Anthropic has declined to change the protocol, so responsibility falls to implementers — treat MCP tool I/O as untrusted and sandbox it. (The Hacker News | The Register | OX Security)
  • If your org runs on Claude Code via the $20 Pro plan, start contingency planning — On Apr 21 Anthropic briefly removed Claude Code from the Pro plan's public docs/landing page before reverting. They confirm an A/B test on ~2% of new prosumer signups. Existing Pro subs aren't affected for now, but leadership is openly signaling subs underprice tokenA 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 ↗ consumption "by a factor of ten." Budget for Max ($100–200/mo) if Claude Code is core to workflow. (The Register | Simon Willison | The New Stack)
  • Read the Bluesky April 6 post-mortem if you run a Go data plane on memcached — A new internal RPC started sending 15–20K URIs per call, each spawning a goroutine and a new memcached TCP conn. Max idle pool = 1000 → TIME_WAIT ephemeral-port exhaustion → synchronous log-on-error → OS-thread explosion → GC pauses → cascade. Fix was group.SetLimit(50) and a per-connection random-loopback dialer. A textbook cautionary tale for anyone running high-fanout RPC to caches behind an LLMLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ service. (Bluesky post-mortem | Surfing Complexity analysis)

Models & Tools

  • ChatGPT Images 2.0 / gpt-image-2 (Apr 21) Trending — OpenAI's first image model with native reasoning: it "thinks" before drawing, renders 2K output, supports 3:1 → 1:3 aspect ratios, and produces up to 8 consistent images from one prompt with character/object continuity. Now available across ChatGPT, Codex, and the API; thinking-enabled outputs are Plus/Pro/Business/Enterprise only. DALL-E 2 and 3 retire May 12. (The New Stack | Axios | PetaPixel)

  • Neura Robotics + AWS collaboration for physical AI (Apr 21) Robotics/IoT Infra — Neura's robotics platform is being linked to AWS cloud/AI infrastructure (SageMaker, Bedrock) to close the real-world training-data gap via simulation-to-operation feedback loops. Notable mainly as a concrete enterprise physical-AI stack to benchmark against NVIDIA's Isaac/GR00T offerings. (The AI Insider)

APIs & Pricing

  • Anthropic Claude Haiku 3 fully retired (Apr 19) Pricing — claude-3-haiku-20240307 now returns errors, not redirects. Migration paths: Haiku 4.5 or Sonnet 4.x tiers. If you still have agents pinned to Haiku 3, they broke this week. (TokenCost migration guide)

  • xAI storage billing live (Apr 20) Pricing — Files and collections stored on the xAI platform now incur storage charges. If you've been using it as a dumping ground during Grok experiments, audit what's still attached. (xAI Docs)

  • Claude Code removed from Pro docs (briefly), A/B test confirmed (Apr 21–22) Pricing Trending — See Action Items. Anthropic reverted the public doc/landing changes but is running the test on ~2% of new prosumer signups. (The Register)

  • Google Gemini billing-tier enforcement reminder (in effect since Apr 1; June 1 hard deprecation looms) Pricing — Gemini 3.1 Pro / 3 Pro / 2.5 Pro removed from free tier; mandatory monthly spend caps ($250 / $2,000 / $20K–100K+). Gemini 2.0 Flash + 2.0 Flash-Lite retire June 1, 2026 — migrate to 2.5 Flash or 3.1 Flash-Lite now. (AI Pricing Guru)

Security

  • Claude Mythos Preview — unauthorized access (Apr 21) Security Trending — Bloomberg, TechCrunch, and Cybernews report that a Discord group obtained access to Mythos on the same day Anthropic announced limited testing, reportedly via URL-pattern guessing and with help from an individual at a third-party contractor. Anthropic says core systems aren't affected; group says intent was curiosity, not harm. CISACybersecurity and Infrastructure Security AgencyThe U.S. cyber defense agency, which works to reduce risk to cyber and physical infrastructure. It runs KEV.CISA ↗ reportedly still has no access. (TechCrunch | Bloomberg | Cybernews)

  • MCP design flaw — RCE in 7K+ servers, 150M+ downloads (Apr 16) Security Open Source — OX Security disclosed a systemic architectural issue in Anthropic's MCP SDKs that enables arbitrary command execution on any system running a vulnerable implementation. Anthropic has declined to modify the protocol, citing the behavior as "expected." CVEs tracking instances include CVE-2026-22252, CVE-2026-22688, CVE-2025-54994, CVE-2025-54136, CVE-2025-49596. (The Hacker News | The Register)

  • CVE-2026-41329 — OpenClaw sandbox bypass (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 ↗ 9.9) (Apr 21) Security Open Source — Critical flaw in OpenClaw (pre-2026.3.31) allows privilege escalationGaining higher permissions than you were given, such as going from a normal user to admin or root.MITRE ATT&CK ↗ via heartbeat context inheritance and senderIsOwner parameter manipulation. Notable because OpenClaw has gone from 9K→210K+ GitHub stars this year and is deployed as a local gateway to integrations like WhatsApp, Slack, Signal, iMessage. Patch immediately. (TheHackerWire)

  • CVE-2026-40933 — MCP Adapters (CVSS 10.0) (this week) Security — Maximum-severity RCE in an MCP adapter due to unsafe serialization of stdio commands; authenticated attackers can inject arbitrary commands into the AI pipeline. Pair with the MCP design-flaw item above when auditing. (SecurityOnline CVE Watchtower)

Industry & Business

  • SpaceX's $60B option on Cursor (Apr 21) Trending — SpaceX announced a partnership plus a right to acquire Cursor for $60B later this year; if SpaceX passes, it pays Cursor $10B for work done. Stack: Cursor's product + SpaceX's "Colossus" supercomputer (~1M-H100 equivalent). Supersedes the Apr 17–19 Thrive/a16z $2B/$50B round. If closed at $60B, this would nearly triple Cursor's November 2025 post-money in six months. (TechCrunch | CNBC | Bloomberg)

  • Project Prometheus nears $10B at $38B (Apr 21) Robotics/IoT — Bezos' five-month-old physical-AI lab (co-CEO with ex-Google X scientist Vikram Bajaj) closes in on a $10B round with JPMorgan and BlackRock; ~120 employees already poached from Meta, DeepMind, SpaceX. Aimed at AI for aerospace and auto manufacturing; Bezos is also reportedly exploring a separate "manufacturing transformation" vehicle up to $100B. (Bloomberg | The Next Web | PYMNTS)

  • Anthropic ↔ Amazon: +$25B, 5GW, $100B AWS commit (Apr 20–21) Infra — Amazon ups its stake by up to $25B (on top of prior $8B); $5B immediately, $20B tied to commercial milestones. Anthropic gets up to 5GW of new compute and commits $100B+ to AWS over 10 years, including Trainium2/3. ~1GW online by end of 2026. Framed explicitly around the recent reliability strain from "sharp rise" in Claude demand. (Anthropic | CNBC | TechCrunch)

AI in Production

  • Bluesky April 6 outage post-mortem (published ~Apr 12) Production Infra — A new internal GetPostRecord RPC began sending 15–20K URIs per call, each spawning a goroutine and its own TCP conn to memcached. Max idle pool of 1000 meant connections got stuck in TIME_WAIT, exhausting ephemeral ports; synchronous error-logging blocked on disk writes, Go spawned ~10x baseline OS threads (~1500 vs ~150), and stop-the-world GC pauses tipped the service into a death spiral. Fix: group.SetLimit(50) + custom dialer that randomizes loopback IPs per conn. Strong artifact for engineers running LLM-adjacent Go services behind memcache or similar. (Post-mortem | Surfing Complexity commentary)

Policy & Safety

  • COPPA AI rule mandatory compliance begins TODAY (Apr 22) Policy — Amended FTC COPPA rules take effect requiring separate parental consent before children's personal information can be used for AI training. First explicit US-law provision covering children's data in AI development. Any team touching consumer products with users under 13 should have privacy/consent flows re-reviewed this week. (Shshell analysis)

  • H.R. 8094 "AI FMTA" remains live in Congress (introduced Mar 26) Policy — Bipartisan Beyer/Lawler/Jacobs bill would require FTC-regulated covered-entity AI developers (10M+ users OR 10²⁶ FLOPs training OR "significant risk") to publicly disclose training data, intended uses, limitations/risks, and eval methods. Fully open-source models are exempt. Enforcement runs through existing FTC powers. Worth tracking because thresholds could shift as markup proceeds. (Congress.gov H.R. 8094 | Alston & Bird | Rep. Beyer press release)

Benchmarks & Evals

  • Gemma 4 on the Arena text leaderboard (released Apr 2; leaderboard climb is current) Open Source — Google DeepMind's 31B Dense variant now sits at #3 on Arena's open-model text leaderboard, with the 26B 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 ↗ at #6. Four sizes (E2B / E4B / 26B MoE / 31B Dense), Apache 2.0, up to 256K context, native vision/audio on smaller tiers. Up to 4x faster and 60% less battery than Gemma 3 on-device. Worth reconsidering as an on-prem / edge baseline if you last benchmarked before April. (Google Blog | Google DeepMind | Hugging Face blog)

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