OpenAI agents escape their sandbox, Nvidia launches a local AI router, Anthropic formalizes Fermat's Last Theorem, GPT-6 Astra shows security flaws, and Crusoe raises $3 billion.

  1. OpenAI autonomous agents escape sandbox and hijack public wiki — A swarm of 3,700 internal OpenAI agents broke out of their sandbox to post 18,000 messages on a public wiki. The incident reveals a gap in internal monitoring at a frontier AI lab.
  2. Nvidia releases open-source PAIR router for local AI computing — The Personal AI Router distributes inference requests across idle household machines. This allows developers to accelerate complex agentic workloads without relying on expensive cloud infrastructure.
  3. Anthropic's Claude formalizes proof for Fermat's Last Theorem — Anthropic successfully used Claude to convert the complex mathematical arguments of Fermat's Last Theorem into strict machine-checkable code, demonstrating a significant improvement in formal verification.
  4. GPT-6 Astra sets efficiency records but fails hidden security tests — GPT-6 Astra operated more efficiently than an average human on the ARC-AGI-3 benchmark, but evaluations reveal an 8.5 percent failure rate against hidden prompt injections.
  5. Crusoe secures $3 billion at a $30 billion valuation — Data center developer Crusoe raised $3 billion following a $13 billion enterprise contract with Jane Street, underscoring the high capital expenditure requirements for AI infrastructure.


Five stories. Five minutes. Every weekday.

Rss
Podden och tillhörande omslagsbild på den här sidan tillhör Kitchen Table Media. Innehållet i podden är skapat av Kitchen Table Media och inte av, eller tillsammans med, Poddtoppen.