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

AI news, Sunday, April 26, 2026

13 stories

TLDR

  • Google to invest up to $40B in Anthropic (Apr 24) — $10B cash up front at a $350B valuation, $30B more on performance milestones; 5 GW of Google Cloud compute over five years. Lands days after Amazon's $25B commitment.
  • Bitwarden CLI supply chain attackAn attack that reaches its target through a supplier, product or service the target relies on.NIST ↗ hits AI coding stacks (Apr 22) — @bitwarden/cli@2026.4.0 was malicious for ~93 minutes; payload explicitly hunts Claude Code, Cursor, Codex CLI, Kiro, and Aider credentials. Treat as confirmed exposure if you installed it.
  • LMDeploy CVE-2026-33626 exploited 13 hours after disclosure (Apr 22) — SSRFServer-side request forgeryA bug where an attacker makes a server send requests for them, reaching internal services it shouldn't expose.OWASP ↗ in the vision-language image loader; Sysdig honeypots saw active port-scanning of cloud metadata, Redis, and MySQL behind the model server.
  • Hugging Face ships ml-intern, an open-source ML engineer agent (Apr 21) — Built on smolagents; lifted Qwen3-1.7B from 10% → 32% on GPQA in under 10 hours and reportedly beat Claude Code's 22.99% on the same task.
  • OpenAI Workspace Agents launch as the GPT successor (Apr 22) — Codex-powered, always-on, shared agents with native Slack/Salesforce/Workspace hooks; free preview through May 6 before credit-based pricing kicks in.

Action Items

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

  • Audit and rotate any credentials touched by @bitwarden/cli@2026.4.0 — The malicious version was live on npm for ~93 minutes on Apr 22 and ran obfuscated JavaScript explicitly designed to exfiltrate developer credentials, cloud secrets, and GitHub Actions secrets — with a payload that targets Claude Code, Cursor, Codex CLI, Kiro, and Aider in particular. CI runners and dev workstations that ran npm install -g @bitwarden/cli@2026.4.0 (or pulled it transitively) should be treated as compromised. Production systems and end-user vault data are unaffected. (The Hacker News | Bitwarden statement | Socket)

  • Patch LMDeploy if you're self-hosting vision-language model inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ — CVE-2026-33626 (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) is an SSRF in the load_image() function of lmdeploy/vl/utils.py — it fetches arbitrary URLs without validating internal/private IPs. Public exploitation began 12h31m after the GitHub advisory: attackers used the VLM as a generic SSRF primitive to scan AWS IMDS, Redis, MySQL, and admin HTTP endpoints. Affects all versions ≤ 0.12.0 with VL support. (The Hacker News | Sysdig | GitLab Advisory)

  • Evaluate OpenAI Workspace Agents before May 6 pricing flip — Codex-powered shared agents are free in research preview for ChatGPT Business/Enterprise/Edu through May 6, then move to credit-based billing. Native Slack/Salesforce/Gmail/Drive integrations and custom MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ server support make this a real GPTs successor; if your org is on a Business+ plan, get a representative agent built and benchmarked while it's free. (OpenAI | VentureBeat)

Models & Tools

  • Hugging Face ml-intern (Apr 21) Open Source Agents — Open-source agent built on smolagents that reads papers from arXiv and HF Papers, sources datasets from the Hub, writes training scripts, runs them on cloud compute, and ships a model. In demos it improved Qwen3-1.7B from 10% to 32% on GPQA in under 10 hours, reportedly outperforming Claude Code's 22.99% on the same task; on a healthcare task it generated 1,100 synthetic samples and upsampled 50× when it judged the existing datasets too low quality. CLI plus desktop/mobile web app. (GitHub | MarkTechPost)

  • OpenAI Workspace Agents (Apr 22) Agents — A successor to custom GPTs: Codex-powered, shared, always-on agents that can be triggered on a schedule or from Slack and run long-running tasks autonomously. Direct integrations for Slack (UI + trigger), Google Workspace (Gmail/Drive/Calendar/Docs/Sheets), and Salesforce out of the box, with custom MCP servers, admin controls, and audit logs. Currently in research preview for ChatGPT Business/Enterprise/Edu/Teachers; free until May 6, then credit-based pricing. (OpenAI | The New Stack | VentureBeat)

Security

  • Bitwarden CLI supply-chain compromise (@bitwarden/cli@2026.4.0) (Apr 22) Production Open Source — A malicious GitHub Action in Bitwarden's CI/CD pipeline (linked to the broader TeamPCP / Checkmarx / "Shai-Hulud" campaign) published a tampered npm package for ~93 minutes. The payload preserves Bitwarden metadata while wiring in an obfuscated JS loader that targets developer credentials, cloud secrets, GitHub Actions secrets, and — notably — credentials/configs for AI coding tools (Claude Code, Cursor, Codex CLI, Kiro, Aider). Browser extensions and the MCP server are unaffected. (The Hacker News | Socket | Endor Labs)

  • LMDeploy SSRF — CVE-2026-33626 (CVSS 7.5) (Apr 22) Open Source Infra — SSRF in load_image() lets attackers point the vision-language image loader at arbitrary URLs, including internal IPs. Sysdig's honeypots logged the first in-the-wild exploitation 12h31m after the GitHub advisory: a single 8-minute session port-scanned AWS IMDS, Redis, MySQL, an admin HTTP interface, and an OOB DNS exfil endpoint, alternating between InternLM-XComposer2 and InternVL2-8B models to dodge anomaly detection. The SSRF-via-LLM-inference attack pattern is generalizable to any tool that fetches user-supplied image URLs. (The Hacker News | Sysdig)

Industry & Business

  • Google to invest up to $40B in Anthropic (Apr 24) Trending Infra — Alphabet committed $10B in cash now at a $350B valuation, with $30B in additional cash + compute conditional on Anthropic hitting performance milestones. Google Cloud will deliver 5 gigawatts of compute to Anthropic over five years (with optional gigawatts on top). Lands four days after Amazon's $25B commitment. Anthropic's annual run-rate revenue is now north of $30B (vs ~$9B at end-2025), driven heavily by Claude Code. (CNBC | Bloomberg | TechCrunch)

  • Meta signs multibillion-dollar AWS Graviton deal — tens of millions of cores for agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ AI (Apr 24) Infra Trending — Meta will deploy hundreds of thousands of AWS Graviton5 chips over a 3+ year window, starting at tens of millions of cores. The bet: agentic AI is CPU-bound (orchestration, real-time reasoning, code generation, search), and Graviton5's 192-core / 5×-larger-cache topology beats x86 on cost-per-orchestration-step. Concurrent with Meta's $115B+ 2026 AI capex. The unusual signal: Meta is a Custom Silicon (MTIA) shop and an Nvidia mega-customer; choosing Amazon's CPU for agentic workloads validates the "CPU is the new GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ for agents" thesis. (CNBC | About Amazon | TechCrunch)

  • Microsoft opens "Rule of 70" voluntary buyout to ~8,750 US employees (Apr 23) Hiring — First voluntary retirement program in Microsoft's 51-year history. Eligibility: senior director and below, age + tenure ≥ 70. Targets ~7% of US workforce; cuts disproportionately fall on long-tenured employees in their 50s–60s — the institutional-knowledge layer most exposed to AI replacement. Comes alongside Microsoft's $80B+ AI infrastructure spend and the company's March 2026 hiring freeze (AI teams excluded). (CNBC | TechCrunch | The Register)

  • Sierra (Bret Taylor) acquires YC-backed Fragment (Apr 23) Agents — Sierra's third tuck-in acquisition of 2026 (after Opera Tech and Receptive AI). Fragment built multilingual conversational AI for customer support — co-founders Olivier Moindrot and Guillaume Genthial join Sierra to lead European agent development. Terms undisclosed; Fragment had raised ~$2M. Continues a rapid wave of consolidation in enterprise AI agent tooling. (TechCrunch)

AI in Production

  • Anthropic publishes "Project Deal" — 186 Claude-on-Claude transactions, $4K total volume (Apr 24) Agents — One-week experiment in Anthropic's SF office where Claude agents bought, sold, and negotiated on behalf of 69 employees in a Slack-based marketplace, with no human-in-the-loop after kickoff. Headline finding: model quality matters and is invisible to users — sellers represented by Opus 4.5 earned $2.68 more per item, Opus-buyers saved $2.45 per item, and Opus-represented users completed 2.07 more deals overall, but participants on Haiku 4.5 reported equal satisfaction. Implication for production agentic systems: outcome-quality A/B tests can't rely on user perception as a proxy for agent skill. (Anthropic | TechCrunch | The Decoder)

  • Character.AI weekend outage rolls into Sunday (Apr 25–26) Production — Character.AI posted multiple "Investigating an issue" status updates spanning hours on Apr 25 covering chats, voices, and character creation; residual problems extended into Sunday Apr 26 with users reporting slow responses, loading errors, and waiting rooms. No formal post-mortem yet. Worth tracking if Character.AI is part of your B2C distribution strategy or you use them as a bellwether for consumer LLMLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ infra. (IBT Australia)

Policy & Safety

  • Florida AG opens criminal investigation into OpenAI over ChatGPT's role in 2025 FSU shooting (Apr 21) Trending — AG James Uthmeier announced subpoenas to OpenAI on Apr 21 after reviewing chat logs from accused FSU gunman Phoenix Ikner, which allegedly include questions about gun/ammo selection, peak times at the FSU student union, and "if there was a shooting at FSU, how would the country react?" Uthmeier's quote: "If it was a person on the other end of that screen, we would be charging them with murder." OpenAI's response: ChatGPT "provided factual responses... that could be found broadly across public sources." First criminal investigation of a major AI lab over a chatbot's role in violence; sets precedent for how prosecutors will frame model outputs as causally linked to user actions. (NPR | CNN | Washington Post)

  • OpenAI launches GPT-5.5 Bio Bug Bounty — $25K for a universal jailbreakA form of prompt injection that makes a model ignore its safety rules.OWASP ↗ (Apr 23) Security — OpenAI is offering up to $25,000 to any researcher who finds a single universal jailbreak that defeats all five questions of its biosafety challenge in Codex Desktop GPT-5.5, from a clean chat with no moderation flag fired. Applications open through Jun 22, testing window Apr 28–Jul 27. NDA-covered. Useful signal: the explicit framing of "universal jailbreak" suggests OpenAI's internal red-teaming has not yet found a single prompt that works against all five questions — a tighter posture than typical jailbreak bounties. (OpenAI | Cybersecurity News)

  • Meta starts capturing employee keystrokes, mouse, and screen content for AI training (Apr 21) Hiring — The "Model Capability Initiative" rolls tracking software onto US employees' work machines to capture clicks, keystrokes, occasional screenshots, and dropdown/keyboard-shortcut usage — explicitly to fix areas where Meta's agents underperform. Meta says the data won't be used for performance evaluations; internal backlash has been notable. The data-source play to watch: as agentic AI hits the "synthetic data is plateauing" wall, employer-controlled telemetry of real workflows becomes the training corpus of choice. (Fortune | TechCrunch)

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