JadidJust-in AI Developments, Incidents & Deadlines
May 2026
SMTWTFS12345678910111213141516171819202122232425262728293031

AI news, Monday, May 25, 2026

11 stories

TLDR

  • Pope Leo XIV's Magnifica Humanitas released today — calls to "disarm" AI, bans on autonomous lethal decisions, presented with Anthropic's Chris Olah (May 25) — Full text published; the Pope says use of AI in warfare "must be subject to the most rigorous ethical constraints" and that entrusting AI with lethal decisions is "not permissible." Already being read in EU/UK policy circles. (Vatican News full presentation | CNN | NCR full text analysis)
  • Anthropic to close ~$30B funding round at $900B+ pre-money valuation as early as this week — surpasses OpenAI (Reported May 22, closing week of May 26) Trending — Sequoia, Dragoneer, Greenoaks, and Altimeter co-lead at ~$2B each; Founders Fund and General Catalyst join. The $900B mark eclipses OpenAI's March $852B. (Bloomberg | Reuters via Yahoo Finance)
  • DeepSeek makes 75% V4-Pro discount permanent — AI 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 ↗ price war escalates (May 23) Pricing — V4-Pro now $0.0036–$0.87 per 1M tokens vs. $0.0145–$3.48 before; the "promo" that was supposed to end May 31 is now the floor. Industry margin compression accelerating. (Bloomberg | Engadget)
  • TrapDoor crypto/AI dev supply-chain attackAn attack that reaches its target through a supplier, product or service the target relies on.NIST ↗ — 34 packages, 384+ versions across npm, PyPI, and Crates.io (Started May 22, active through weekend) Security — Plants malicious .cursorrules and CLAUDE.md files designed to trick Cursor and Claude Code into running a fake "security scan" that exfiltrates secrets. Targets AI dev tooling specifically. (Socket research | The Hacker News)
  • Anthropic in early talks to rent Microsoft Maia 200 AI chips — would be the first external Maia customer (May 21) Infra — Adds a fifth silicon vendor (NVIDIA / AWS Trainium / Google TPUTensor Processing UnitGoogle's custom chip built to speed up machine learning, especially large matrix operations.Google Cloud ↗ / SpaceX / Microsoft) to Anthropic's multi-supplier strategy. Microsoft's Maia 200 claims ~30% better tokens/$ vs. prior silicon. (CNBC)

Action Items

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

  • GitHub Copilot moves to usage-based billing in 7 days (June 1) — Plan list prices don't change, but anyone on an annual Pro/Pro+ plan staying on request-based billing will see model multipliers jump sharply (Opus 4.7 goes from 7.5x to 27x). Move now to confirm: (1) which seats are annual vs. monthly, (2) which models your team actually uses (Chat / Copilot CLI / cloud agent / Spaces all consume credits; completions stay unlimited), and (3) whether code review's GitHub Actions minute consumption changes your CI cost profile. Budget-impact teams should reforecast before June. (GitHub blog | GitHub Docs: usage-based billing prep | Annual plan multipliers table)

  • Scan repos and dev machines for TrapDoor's AI-assistant trojan files — Beyond the usual lockfile audit for the 34 compromised packages on npm/PyPI/Crates.io, grep every cloned repo for .cursorrules and CLAUDE.md files containing instructions to run a "security scan," "credentials check," or any command that touches ~/.ssh, ~/.aws, wallet keystores, or env files. The campaign is novel in weaponizing AI-IDE rule files; standard npm audit and Dependabot will miss it. Detection time from Socket is < 6 minutes per package but the rule-file payloads sit in cloned repos indefinitely. (Socket TrapDoor breakdown | The Hacker News)

  • If you ship a GPAI model into the EU, you have ~10 weeks until full Commission enforcement (August 2) — Article 99 fines and the Commission's full investigative powers (documentation requests, evaluations, market restriction, recall, withdrawal) all become operational on August 2, 2026. Models released before August 2, 2025 get an extra year (until August 2, 2027) but post-cutoff models have no grace. Confirm: (1) you've registered the model, (2) Code of Practice signatory status (or your alternative compliance evidence), (3) downstream-provider documentation kit, and (4) a named EU representative. (European Commission AI Act page | AI Act Chapter V enforcement explainer)

Models & Tools

  • Google Antigravity GA + Gemini 3.5 Flash as the new agent-default Flash model (May 19, I/O 2026) Agents — Gemini 3.5 Flash launched at Google I/O 2026 (May 19–20) as the new default in AI Mode and the production model behind Google Antigravity, Google's agent-first dev platform. Google claims it outperforms Gemini 3.1 Pro on coding and agent benchmarks at Flash latencies. Available via the Gemini API in AI Studio, Android Studio, and Antigravity. (Google blog: 100 things announced at I/O 2026 | Sundar Pichai I/O 2026 keynote post)

  • Gemini Spark — Google's 24/7 personal agent — opens MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ support this week, AI Ultra beta lands next week in the US (May 19 announce; rollout May 25–early June) Agents — Built on Gemini 3.5 Flash + the Antigravity harness. Day-one MCP connections include Canva, OpenTable, and Instacart; Slack, Salesforce, Notion, and custom MCP servers expected in the coming weeks. Trusted-tester wave is now; Beta to Google AI Ultra subscribers in the US follows. Read it as Google making MCP a first-class user-facing surface, not just a dev primitive. (Google blog: Gemini app becomes more agentic)

  • OpenAI Codex "Locked Use" on macOS — phone-triggered, runs while your Mac is locked (May 21) Agents — A new Computer Use plugin uses an Apple authorization plug-in to temporarily unlock your Mac, click through apps, and run long-horizon tasks while the screen is off. Requires Screen Recording + Accessibility permissions per app; Terminal, Codex itself, and admin prompts are blocked. Unavailable in EU/UK/Switzerland at launch. Closes the "laptop sitting open on the desk" gap for remote agent work. (MacRumors | Codex changelog | Macworld)

  • OpenAI + Dell — Codex on hybrid/on-prem via Dell AI Data Platform and AI Factory (May 18) Infra — OpenAI's first major hybrid/on-prem distribution deal: Codex agents will run against data already governed by Dell AI Data Platform; the wider scope explores ChatGPT Enterprise + API solutions running inside Dell AI Factory. Codex now claims >4M weekly developers; this is the route to regulated-industry adoption that doesn't require shipping data to OpenAI's cloud. (OpenAI announcement)

APIs & Pricing

  • DeepSeek: V4-Pro 75% discount made permanent (May 23) Trending — See TLDR. The price page now lists V4-Pro at $0.0036–$0.87 per 1M tokens as the standing rate after May 31. Combined with Anthropic's enterprise discounting and Gemini Flash GA pricing pressure, this confirms the "high-margin AI tokens" era is closing. (Bloomberg | Engadget | The Next Web)

  • GitHub Copilot: usage-based billing + AI Credits live June 1 (Effective June 1, 2026) Pricing — See Action Items. 1 AI Credit = $0.01; each paid plan includes a monthly allotment equal to its dollar price ($10 Pro / $19 Business / $39 Pro+/Enterprise). Code completions and next-edit suggestions remain unlimited; Chat, CLI, cloud agent, Spaces, Spark, and third-party coding agents draw down credits. (GitHub blog)

Security

  • TrapDoor — multi-ecosystem supply-chain attack weaponizes Cursor / Claude rule files (First sighting May 22 20:20 UTC; active through weekend) Trending Open Source — Socket is tracking 34+ malicious packages across 384+ versions on npm, PyPI, and Crates.io. The novelty: alongside standard postinstall / build.rs / Python import-time payloads, packages plant .cursorrules and CLAUDE.md files with hidden instructions that nudge Cursor and Claude Code into running a fake "security scan" — the scan exfiltrates SSH keys, AWS creds, GitHub tokens, browser logins, env files, and crypto wallets (Sui, Solana, Aptos). Mean detection by Socket was under 6 minutes, but the rule-file persistence in cloned repos is the longer-tail risk. Targets cluster around AI-assistant users, Solana/DeFi, and Rust dev tooling. (Socket research | The Hacker News | Cybersecurity News)

Industry & Business

  • Anthropic: $30B round at $900B+ pre-money closing as soon as this week (Reported May 22; expected close week of May 26) Trending — See TLDR. Sequoia, Dragoneer, Greenoaks, and Altimeter co-lead at roughly $2B each; Peter Thiel's Founders Fund and General Catalyst participate among existing backers. Closes a fortnight that also brought Q2 revenue guidance of $10.9B and Anthropic's first quarterly operating profit. Surpassing OpenAI on private valuation reshapes the talent/compute leverage Anthropic can pull during the back half of 2026. (Bloomberg | Reuters via Yahoo Finance | FT via Investing.com)

  • Anthropic ↔ Microsoft: early talks for Anthropic to rent Azure Maia 200 capacity (May 21) Infra — See TLDR. No deal signed, both companies declining to comment per CNBC. If it closes, Maia gets its first external frontier-model customer — meaningful for Microsoft, which has been a step behind AWS Trainium and Google TPU on the custom-silicon-as-a-business front. (CNBC | Cloud Computing News)

Policy & Safety

  • Pope Leo XIV — Magnifica Humanitas — first papal encyclical on AI, released May 25 (May 25) Trending — See TLDR. The document explicitly calls for AI to be "disarmed," names autonomous weapons as a category requiring regulation, says lethal decisions are "not permissible" to delegate to AI, and warns that the growing ease of deploying autonomous weapons systems makes war "more feasible and less subject to human control." Cardinal Parolin closed the launch with concrete asks for international regulation. Watch for the document to appear in EU AI Act and UK AISI floor speeches within weeks; it's already cited by Vatican Press Office as the basis for the new Vatican AI commission's mandate. (Vatican News presentation summary | Full-text coverage — National Catholic Register | NBC News | US News & World Report)

  • EU AI Act: Commission enforcement powers against GPAI providers go live August 2 (10 weeks out) Policy — See Action Items. From August 2, 2026 the AI Office can demand documentation, run independent evaluations, order risk-mitigation measures, restrict market access, recall systems, and levy fines (up to €15M or 3% of worldwide annual turnover for GPAI breaches). Providers of GPAI models placed on the EU market after August 2, 2025 must already be compliant; pre-cutoff models have until August 2, 2027. (European Commission | Chapter V enforcement explainer)

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