OpenAI debuts ChatGPT Images 2.0 as Anthropic breach, SpaceX‑Cursor Colossus deal, Moonshot agents and Meta keystroke tracking reshape AI

AI Tech News Today — Aurora & Isabelle
Date: 2026-04-22T12:11:27.000Z
Salute.

Five fast hits on AI developments that matter this morning:

1. OpenAI launches ChatGPT Images two point zero

• What’s new: OpenAI rolled out ChatGPT Images two point zero, powered by GPT-Image-two, a model that “thinks” through layouts before it draws.
• Key capabilities:
  • Plans complex infographics and consistent image campaigns.
  • Produces up to eight matching two K-resolution images per prompt.
  • Handles tricky typography far better than before.
• Developer tools: New features include a Chronicle memory and an open-sourced visualizer called Euphony to track sessions.

2. Anthropic’s Mythos model briefly accessed by unauthorized users

• Report: Bloomberg says Anthropic’s unreleased Mythos cyber model was reached through a third-party environment.
• Company response: Anthropic now requires government-issued ID and selfies for sensitive accounts.
• Broader moves: Anthropic is doubling down on persistent AI with Conway, investing in always-on agents and striking a large cloud deal for five gigawatts of capacity and more than one million custom Trainium two chips.

3. SpaceX partners with Cursor to train elite models on Colossus

• Partnership: SpaceX gave coding startup Cursor access to Colossus — a supercomputer equivalent to about one million H one hundred GPUs.
• Commercial terms: SpaceX has an option to acquire Cursor for sixty billion dollars later this year or accept a ten billion dollar development fee.
• Cursor’s goal: Train what it calls the world’s best coding and knowledge-work models on that hardware.

4. Moonshot AI’s Kimi K two point six runs autonomous agents for five days

• Milestone: Moonshot’s Kimi K two point six kept an agent swarm running across three hundred sub-agents and four thousand coordinated steps, completing a compiler in ten hours (work that often takes months for teams).
• Implication: Shows persistent agents can perform heavy lifts — but also highlights risks around brittle orchestration and the urgent need for agent gateways and kill switches.

5. Meta is tracking employee keystrokes to train AI

• Report: Meta is reportedly logging employee keystrokes to improve its models.
• Concerns: Raises familiar privacy and consent questions and underscores a broader industry trend of pulling more internal signals into model training pipelines.

Closing
That’s the roundup from Aurora and Isabelle — quick, direct, and hopefully useful. Stay curious, keep an eye on persistent agents and security, and we’ll be back with more AI news soon.