Pentagon bans Anthropic + OpenAI's browser leaked

Databricks AI Research

Databricks unveils KARL, an RL-trained enterprise search agent that outperforms Claude 4.6 and GPT 5.2 across diverse search tasks using multi-task training and synthetic data generation. The system demonstrates superior cost-efficiency and generalization, including on out-of-distribution tasks.

MIT, Meta, Google DeepMind, and Google Research

LLMs trained to mimic a Bayesian reasoning model through "Bayesian teaching" significantly outperform both off-the-shelf models and oracle-trained versions in tasks requiring probabilistic belief updates. This distillation approach enables LLMs to generalize probabilistic reasoning skills to new domains, including real-world applications where implementing exact Bayesian models is impractical.

The Pentagon has designated Anthropic a national security supply chain risk, banning military contractors from working with the company over disputes regarding unrestricted military access to Claude AI models. The move includes a federal prohibition on Anthropic's technology with a six-month transition period.

BleepingComputer

OpenAI is building a Chromium-based browser that will allow ChatGPT Agent to directly control local browser functions, moving beyond its current cloud-only virtual browsing capabilities. The leaked code reveals a toggle between cloud and local browser execution, suggesting OpenAI aims to consolidate more user interactions within its own ecosystem.

X (formerly Twitter)

OpenAI launched a Codex CLI tool with GPT-5 support and configurable reasoning effort, available to ChatGPT subscribers through npm installation.