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

AI news, Saturday, April 25, 2026

10 stories

TLDR

  • Cohere acquires Aleph Alpha (Apr 24) — Canada's enterprise AI lab buys Germany's AI pioneer in a sovereign-AI play; $600M Series E from Schwarz Group; combined entity valued at ~$20B.
  • Tesla raises 2026 capex to $25B+ (Apr 22) — Nearly 3× last year's spend; Optimus humanoid factories, robotaxi infrastructure, and a joint Terafab chip fab with SpaceX are the targets.
  • Google Antigravity patched for sandbox-escape RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ (Apr 21) — Prompt injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ in the file-search tool bypasses Secure Mode and enables full code execution; update now if you use it.
  • Claude Opus 4.7 launched (Apr 16) — Now #1 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 (87.6%); adds task budgets for agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ loops and hi-res image support; price held at $5/$25 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 ↗.
  • Arizona AI bills hit final adjournment day (Apr 25) — Three bills racing the legislature's April 25 deadline; mandatory agency AI-adoption bill is closest to passage.

Action Items

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

  • Update Google Antigravity if you use it — A sandbox-escape RCE was publicly disclosed April 21: prompt injection in the find_by_name tool bypasses Secure Mode entirely, enabling arbitrary code execution before security restrictions are evaluated. Google patched the issue; confirm your org is on the latest version and treat unvalidated file-path inputs in any agent pipeline as attacker-controlled. (Dark Reading | Pillar Security writeup)

  • Benchmark Claude Opus 4.7 for your agentic workloads — Now #1 on SWE-bench Verified (87.6%) and SWE-bench Pro (64.3%), shipping since April 16 with task budgets and 2576px image support. Price held at $5/$25 per M tokens while GPT-5.5 doubled to $5/$30 — the cost argument for switching to Opus 4.7 for long-horizon agent loops is material. (Anthropic)

  • Evaluate Databricks Unity AI Gateway for MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ governance — Now GA inside Unity Catalog; enforces fine-grained MCP server permissions (on-behalf-of access), PII redaction, prompt injection detection, and per-team LLMLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ cost attribution across agent-to-MCP calls. Directly relevant if you're building multi-agent pipelines on Databricks. (Databricks Blog)

Models & Tools

  • Claude Opus 4.7 (Apr 16) Agents — Anthropic's new flagship adds task budgets (token-target hints that help agents self-regulate loop depth), high-resolution image input (up to 2576px / 3.75MP), and improved tool-use reliability in long workflows. Tops SWE-bench Verified at 87.6% and SWE-bench Pro at 64.3% — both #1 among generally available models. Available day one on the API, Bedrock, Vertex, Foundry, and GitHub Copilot; price unchanged at $5/$25 per M tokens (same as Opus 4.6). (Anthropic | GitHub Changelog)

  • Databricks Unity AI Gateway (Apr 15) Infra Agents — Databricks' AI Gateway is now part of Unity Catalog, extending its governance model to agentic AI. Key capabilities: fine-grained MCP server permissions, PII detection/redaction, prompt injection detection, data exfiltration prevention, hallucination guard, and per-team LLM cost attribution. Single unified API across Claude, GPT, Gemini, Llama, and any provider-native endpoint, with built-in fallbacks and rate limits. (Databricks Blog | AWS Docs)

  • NVIDIA Ising (Apr 14) Infra — NVIDIA's first open AI model family for quantum computing: a 35B vision-language model (Ising Calibration) that automates qubit calibration from days to hours, and a 3D CNN decoder (Ising Decoding) for real-time quantum error correction — claimed 2.5× faster and 3× more accurate than traditional methods. Open-source; early adopters include Harvard, Fermilab, and Lawrence Berkeley. Not directly in your CI pipeline today, but the AI+quantum intersection is maturing faster than expected. (NVIDIA Newsroom | CNBC)

Security

  • Google Antigravity sandbox-escape RCE via prompt injection (Apr 21) Agents — Pillar Security disclosed that Antigravity's find_by_name file-search tool insufficiently sanitizes its Pattern parameter, allowing attackers to inject flags into the underlying fd binary — converting a file lookup into arbitrary code execution. The critical detail: the vulnerable tool call fires before Secure Mode restrictions are evaluated, making Secure Mode irrelevant as a mitigation. Google patched the issue and awarded a $10K VRP bounty. General lesson: AI agent tool parameters that touch the filesystem or shell are high-value injection targets regardless of declared security modes. (Dark Reading | CyberScoop | Pillar Security)

Industry & Business

  • Cohere acquires Aleph Alpha; Schwarz Group invests $600M (Apr 24) Trending — Canadian enterprise AI lab Cohere agreed to buy German AI startup Aleph Alpha (which pivoted away from frontier model development in late 2025 to focus on specialized enterprise applications) in a sovereign-AI play targeting Europe's regulated sectors: public sector, finance, defense, and healthcare. The deal comes alongside a $600M Series E from Schwarz Group (Europe's largest retailer, an existing Aleph Alpha backer). Combined entity valued at ~$20B with global HQ in Toronto and European HQ in Berlin under Cohere CEO Aidan Gomez; Cohere shareholders receive ~90% of the combined entity. Regulatory approval pending. (CNBC | BusinessWire)

  • Tesla raises 2026 capex to $25B+ — nearly 3× last year's $8.53B (Apr 22) Robotics/IoT — Targets: a dedicated Optimus humanoid robot manufacturing facility outside Austin, robotaxi infrastructure, and Terafab — a joint chip fab with SpaceX for in-house AI silicon. Musk framed Tesla as transitioning "from a hardware-centric business to a physical AI company." The production of Model S and Model X will wind down at Fremont to make room for Optimus at scale. (Bloomberg | TechCrunch)

  • Instacart co-founder Apoorva Mehta launches Abundance, a fully AI-run hedge fund (Apr 24) Agents — $100M seed; AI agents autonomously source trade ideas, conduct research, select long/short positions, size bets, and execute trades. Mehta attributes the inspiration to OpenAI's o3 reasoning modelA model that works through hidden reasoning steps before its final answer. Those steps are billed as output.OpenAI ↗. Fund currently operates on proprietary capital with plans to open to external investors. Whether AI agents can sustain alpha is unproven — but the capital and pedigree are real. (Bloomberg)

Policy & Safety

  • Arizona AI bills at adjournment deadline today (Apr 25) — Three bills racing the legislature's April 25 adjournment: (1) HB 2592 — mandates all state agencies identify AI adoption opportunities and remove regulatory barriers; passed both chambers, in reconciliation; (2) HB 2311 — kids chatbot safety bill; bipartisan support but stalled in Senate; (3) HB 2371 — AI-assisted arbitration in divorce proceedings; stalled since March 30. The pattern worth noting: mandatory AI-adoption mandates are advancing while consumer protection bills are stalling — the inverse of what most EU-watching teams expect from state legislatures. (AZ Mirror | Transparency Coalition)

  • EU AI Act Phase Two — 99 days out (Aug 2, 2026) — New transparency requirements and high-risk AI system compliance rules take effect for all companies operating in Europe. If you're shipping AI features that touch EU customers and haven't started your high-risk system assessment, the runway is now under 100 days. (Alston & Bird AI Quarterly)

Benchmarks & Evals

  • SWE-bench Verified April 2026 leaderboard — Claude Opus 4.7 leads generally available models at 87.6% (SWE-bench Verified) and 64.3% (SWE-bench Pro), both #1. Claude Mythos Preview tops the raw leaderboard at 93.9% but relies on non-public scaffolding. GPT-5.3-Codex follows at 85.0% Verified; Claude Opus 4.5/4.6 and Gemini 3.1 Pro cluster at 80–81%. The gap between Opus 4.7 and the next tier (~2.6 points) has widened materially over the past two weeks. (SWE-bench.com | Marc0 Leaderboard)

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