AI Tech News Today — Briefing

Hello — salute from Aurora and Isabelle at AI Tech News Today.

Date: 2026-06-05T15:47:57.000Z

Anthropic pauses testing after Oceanus leak

• What happened: Anthropic says it paused safety testing of a secret Mythos model, codenamed Oceanus, after an unreleased build leaked through third‑party proxies.
• Internal impact: The company warns about recursive self‑improvement: Claude now writes over 80% of Anthropic’s production code, and developer output has reportedly jumped eightfold.
• Security concerns: Reports say the National Security Agency has forward‑deployed about half a dozen Anthropic engineers to help deploy Claude Mythos for cyber operations — raising concerns about offensive uses.
• Commercial moves: Anthropic is launching a Partner Hub and a new command‑line tool called ant to help partners build with Claude — a reminder that safety debates and commercial rollout are happening at the same time.

OpenAI rolls out Dreaming memory and life‑sciences upgrades

• Dreaming (background memory): OpenAI announced Dreaming, a background memory system that:
  • quietly builds structured dossiers from your chat history,
  • adapts as your context changes, and
  • reportedly cuts backend compute costs by 5×.
    Dreaming is slated to reach free tiers.
• Life‑sciences model: OpenAI upgraded GPT‑Rosalind to handle complex genomics and drug‑discovery workflows while using about 31% fewer tokens.
• Policy & outreach: CEO Sam Altman is pushing a voluntary 30‑day model‑sharing idea with government and is publicly distancing the company from political funding — part of a charm offensive in Washington.

Google ships Gemma for laptops — doubles down on local first

• Gemma 4 12B: Google DeepMind introduced Gemma 4 12B, an Apache‑2 licensed multimodal model designed to run on ordinary laptops with 16 GB of memory.
  • Key design: an encoder‑free architecture that feeds vision and audio directly into the decoder.
  • Context capability: supports a 256,000‑token context window.
• Local tools: Google pairs Gemma with local creative tools like Magenta RealTime 2 and an experimental Dreambeans app for paid AI subscribers as the company pivots toward on‑device models and self‑improvement loops.

Security: AI accelerates both defense and attack

• Faster discovery than patching: Anthropic’s Logan Graham warns that AI is finding flaws faster than teams can patch them.
• Adaptive AI malware: New research shows AI malware that reads live vulnerability advisories and adapts attacks in real time.
• HTTP/2 bomb DoS: A fresh HTTP/2 “bomb” denial‑of‑service can crash web servers in under a minute.
• Biometric privacy: Reporting that Meta quietly shipped three live biometric models inside phone software has privacy watchers on alert.

That’s it for today. Stay curious, stay skeptical, and we’ll keep listening so you don’t have to.