1. Apple to Tighten macOS Full Disk Access, Requiring “Very Explicit” User Action for AI Agents Apple is tightening one of macOS’s most powerful permissions as AI agents gain broader access to personal data.

2. Cloudflare Releases Open-Weight Clef Decision Models and Launches Reinforcement-Learning Fine-Tuning Platform Cloudflare has released Clef and Clef-flash, two models designed to make bounded decisions inside automated workflows.

3. Ataraxos Beats a Top Stratego Player 15-1 After Training on Just 16 GPUs An AI system called Ataraxos defeated top Stratego player Pim Niemeijer 15 games to one, with four draws, overcoming a classic game that had resisted systems capable of mastering chess, Go, and poker.

In Brief

• Anthropic Commits $100 Million to Train 10,000 Enterprise AI Engineers Anthropic launched Claude Frontier Academy, aiming to train 10,000 “Frontier Deployed Engineers” by the end of 2027. Initial cohorts include engineers from Accenture, Deloitte, Morgan Stanley and other large organizations.

• DeepSeek Releases an Open-Source Desktop Agent Harness DeepSeek Harness, now in preview for macOS and Windows, lets users install or create plugins for tasks including document organization, spreadsheet analysis, coding and scheduled work. Developers can inspect execution traces, run its web interface through Node.js or clone the source from GitHub.

• OpenAI Publishes a Developer Guide to the GPT-6 Model Family OpenAI released guidance for startups on selecting GPT-6 models, configuring reasoning effort, improving prompts and skills, coordinating tools and preparing workflows for production.

• Chatham Financial Cuts Trade Validation From 30 Minutes to Under Four Chatham Financial, a capital-markets advisory and technology firm, is using Codex and GPT-5.6 to build software and redesign workflows. OpenAI says the work reduced trade-validation time from 30 minutes to less than four.

• Explicit Belief States Improve Long-Horizon Agent Performance in Tests Researchers introduced Progression of States, an inference-time framework that maintains an agent’s current-world beliefs, detects stalled progress and applies targeted recovery without additional training. The authors report the best overall performance across four benchmarks and all three tested LLM backbones.

• OneStreamer Gives a 4B Video Model Persistent Memory and Proactive Responses Researchers introduced OneStreamer, which records timestamped captions and event summaries from live video so it can answer later questions without revisiting older visual features. Its authors report the strongest results among compared methods across eight streaming-video benchmarks and released the OneStreamer-1M dataset with more than one million records.

Read the full edition →