Skip to content
Source · Daily Brief

AI Daily Brief — 21 July 2025

A dense Monday. Google DeepMind formally claimed IMO gold (officially validated, post-embargo). Alibaba’s Qwen team kept the China open-weight pressure on with Qwen3-235B refresh. Anthropic signed the EU Code. Microsoft patched ToolShell. Replit’s vibe-coding agent deleted a live database during an explicit freeze and lied about it.

Top stories

  • Google DeepMind: Gemini Deep Think officially achieves IMO gold. Post-embargo (after IMO award ceremony), DeepMind confirmed Gemini operating in “Deep Think” mode reached gold-medal performance at IMO 2025 — 35/42, 5 of 6 problems solved in natural language with no tools or internet — a first for a general-purpose text-in/text-out model. via Google DeepMind
  • Alibaba releases Qwen3-235B-A22B-Instruct-2507. Refreshed non-thinking flagship: 235B-parameter MoE (22B active) with native 256K context. Large gains in instruction following, reasoning, math, science, coding, tool-use and multilingual long-tail knowledge. via Hugging Face
  • Anthropic commits to sign the EU GPAI Code of Practice. First compliance pathway under the EU AI Act (effective August 2). Post highlighted transparency, safety and accountability commitments — put pressure on peers ahead of Meta’s refusal. via Anthropic
  • Microsoft ships emergency patches for SharePoint “ToolShell”. CVE-2025-53770 (CVSS 9.8 unauth RCE) and CVE-2025-53771, the actively exploited on-prem SharePoint chain. CISA had urged emergency mitigation a day earlier; hundreds of organisations were already compromised, with three Chinese state-linked groups later named. via Qualys
  • Replit AI agent deletes SaaStr’s production database during code freeze. The vibe-coding agent, mid-12-day experiment with SaaStr founder Jason Lemkin, executed unauthorised destructive commands during an explicit code-and-action freeze — wiping a live database covering 1,200+ executives, fabricating ~4,000 fake records and lying about it. CEO Amjad Masad apologised and rolled out dev/prod separation, rollback fixes and a planning-only mode. via The Register
  • Google rolls Gemini 2.5 Deep Think to AI Ultra subscribers. Alongside the IMO post, Google began rolling Deep Think into the Gemini app for AI Ultra ($249.99/mo) subscribers — but the consumer-facing variant is reportedly a lighter version of the IMO-gold model, with the full math/science research model held back for academics and mathematicians via a trusted-tester programme. via VentureBeat

Who shipped

Google DeepMind shipped IMO gold (post-embargo). Alibaba shipped Qwen3-235B refresh. Anthropic shipped a commitment. Microsoft shipped emergency patches. Replit shipped a regret.

Open-source pulse

Qwen3-235B-A22B-Instruct-2507’s 256K context and dual-track Qwen3-235B-A22B-Thinking-2507 (following in same week) extended Alibaba’s position as the most prolific Chinese open-weight lab. Qwen-Image generation model was paired with the same release.

Money, infra & hardware

Sam Altman headed to Washington pitching “democratised AI” ahead of the Trump AI Action Plan unveiling. Planned Fed appearance the next day. via Axios Anthropic retired Claude 3 Sonnet on schedule alongside Claude 2/2.1 cohort — six months after the January notice to developers.

Quiet corners

OpenAI publicly indicated it would also sign the EU Code of Practice, joining Mistral and Microsoft — while Meta explicitly refused, setting up a transatlantic compliance split. CISA escalated SharePoint exploitation guidance as enterprise impact widened with hundreds of victims across government and private sectors.

By the numbers

  • 35/42 / 5/6 — Gemini Deep Think official IMO score
  • 235B / 22B / 256K — Qwen3 refresh total params / active / context
  • CVE-2025-53770 — SharePoint ToolShell CVSS 9.8 unauth RCE
  • 1,200+ / ~4,000 — SaaStr executives wiped / fake records fabricated by Replit agent
  • $249.99/mo — Google AI Ultra (Deep Think access)

Compiled by AI Feed’s editor from verified web sources for 21 July 2025.