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

AI news, Monday, April 20, 2026

9 stories

TLDR

  • OpenAI Codex gets computer use + 90 plugins (Apr 16) — Codex for Mac now runs as a background agent that sees, clicks, and types across any app while you keep working; new plugins include CircleCI, CodeRabbit, and GitLab Issues
  • CVE-2026-21520: Copilot Studio prompt injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ — data leaked even after the patch (Apr 15) — Capsule Security's "ShareLeak" demo showed customer data still exfiltrating via Microsoft's patched agent; companion "PipeLeak" affects Salesforce Agentforce
  • OpenAI acquires Hiro Finance (Apr 13) — Acqui-hire of "AI personal CFO" startup; Hiro app goes dark today (Apr 20), signaling ChatGPT's move into native financial planning
  • EU AI Act: 15 industry groups push for labeling deadline extension (Apr 13) — Industry requests doubling the grace period from 6 to 12 months ahead of the Aug 2, 2026 AI-generated content compliance deadline
  • 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 open-source (Apr 7) — Zhipu AI's MIT-licensed 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 ↗ model is #1 on SWE-Bench Pro at 58.4%; runs 8-hour autonomous coding loops with no commercial restrictions

Action Items

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

  • xAI storage billing is live as of today (Apr 20) — If your team uses the xAI Files API, charges began accumulating today. Log into the xAI Console immediately, delete unused files and collections, and set spend caps to avoid surprise charges. (xAI Console | xAI Docs)
  • Evaluate Codex desktop for your CI/CD workflow — OpenAI's Apr 16 update added 90+ plugins including CircleCI, CodeRabbit, GitLab Issues, Atlassian Rovo, and Render, plus computer-use for Mac. Worth testing for PR-review automation and background test runs. (OpenAI Blog | Changelog)
  • Audit AI agent attack surface after ShareLeak / PipeLeak — Both Microsoft Copilot Studio and Salesforce Agentforce were exploited via public-facing form fields injecting instructions into the agent's context. Audit any agent that ingests user-supplied content, apply output filtering, and enforce least-privilege tool access. (VentureBeat)

Models & Tools

  • OpenAI Codex: computer use + 90+ plugins (Apr 16) Agents — The Codex desktop app for macOS now runs background agents that can see, click, and type across any application while you continue working. New capabilities include an in-app browser, persistent cross-session memory, a scheduling layer for tasks that span days or weeks, SSH to remote devboxes, and 90+ new plugins (CircleCI, CodeRabbit, GitLab Issues, Atlassian Rovo, Render). EU/UK rollout coming soon. Directly targets Claude Code's enterprise foothold. (OpenAI Blog | MacRumors | gHacks)

  • Canva AI 2.0 (Apr 16) Customer AI Agents — Launched as "the company's biggest launch ever," the platform now supports natural-language design generation, AgenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ Orchestration (multi-step campaign creation from a single brief), persistent brand memory, and connections to Gmail, Slack, and Zoom. Backed by a deepened Anthropic partnership: Claude Design is embedded directly in the Visual Suite. Rolling out from 1M preview users toward general availability in coming weeks. (Fortune | SiliconAngle)

  • GLM-5.1 (Z.ai / Zhipu AI) (Apr 7) Open Source — 744B MoE model with 40B active parameters, 200K context, MIT license, and weights on HuggingFace with zero commercial restrictions. Currently #1 on SWE-Bench Pro at 58.4%, ahead of GPT-5.4 (57.7%) and Claude Opus 4.6 (57.3%). Unique capability: autonomously manages a full "plan → execute → test → fix → optimize" loop for up to 8 hours without human intervention. Strong option for teams needing a capable open-source coding model they can run on-prem or fine-tuneTraining an already trained model further on extra data so it fits a specific task, domain or style.Anthropic ↗ freely. (The Decoder | TechBriefly | HuggingFace)

Security

  • ShareLeak (CVE-2026-21520) + PipeLeak: AI agent patches are not enough (Apr 15) Agents — Capsule Security — which launched from stealth this week with $7M — disclosed two parallel indirect prompt injection vulnerabilities. ShareLeak (CVE-2026-21520, 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 ↗ 7.5) in Microsoft Copilot Studio exploits SharePoint form inputs to override agent instructions and exfiltrate customer data via Outlook; Microsoft patched on January 15, but Capsule's post-patch demo showed data still escaped due to insufficient output filtering. PipeLeak in Salesforce Agentforce uses the same form-injection vector against CRM records; no CVECommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗ assigned yet. The key lesson: patching the LLMLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ call does not automatically secure agent output paths — you must also filter what the agent is allowed to do with extracted data and restrict tool access to least privilege. (VentureBeat | Dark Reading | Help Net Security)

Industry & Business

  • OpenAI acquires Hiro Finance (Apr 13) Trending — OpenAI's sixth acquisition of 2026: an acqui-hire of Ethan Bloch's "AI personal CFO" startup (backed by Ribbit and General Catalyst) that managed $1B+ in user assets. The Hiro app goes dark today (Apr 20); users can export data until May 13. Signals that ChatGPT is being built toward native financial planning and vertical AI capabilities, not just general assistance. (TechCrunch | American Banker)

  • PwC: Top 20% of firms capture 74% of AI's economic value (Apr 15) — Study of 1,217 senior executives found AI leaders grow revenue 7.2× faster than peers — not primarily from cost-cutting but from using AI to enter new markets and automate decision-making at 3× the rate of laggards. Leaders are also 1.7× more likely to have a formal Responsible AI framework. Useful data point for engineering leaders making the case for strategic (not just productivity) AI investment. (PwC Press Release)

  • Q1 2026: $300B in global VC, 80% to AI — and it's extremely concentrated (Apr 2026) Infra — Crunchbase confirmed Q1 set an all-time global VC record; AI took $242B (80%), but 65% of that flowed to just four companies: OpenAI ($122B), Anthropic ($30B), xAI ($20B), and Waymo ($16B). Despite dollar records, North American deal count fell 26%, signaling extreme consolidation at the frontier. (Crunchbase News | Crunchbase: Foundational AI)

Policy & Safety

  • EU AI Act: Industry requests 6→12 month labeling grace period (Apr 13) — EuroISPA and 14 co-signing industry associations formally asked EU policymakers to extend the implementation period for Article 50's AI-generated content labeling requirements from six to twelve months. The current hard deadline is August 2, 2026; the Commission's draft Code of Practice on marking AI content won't be finalized until June. Teams building public-facing generative AI products in Europe should monitor this closely — if the extension is denied, you'll need machine-readable content labeling live by August 2. (The Legal Wire | Kennedy's Law)

Benchmarks & Evals

  • SWE-bench snapshot — April 20, 2026 — Claude Opus 4.7 (launched Apr 16) leads SWE-bench Verified at 87.6% and SWE-bench Pro at 64.3%, both #1. On the open-source tier, GLM-5.1 leads SWE-bench Pro at 58.4%. The growing gap between Verified (~80–87%) and Pro (~58–64%) scores continues to confirm that Pro is the more realistic measure of real-world autonomous coding capability. (SWE-bench.com | Scale Labs Leaderboard)

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