Mollick: steering long-running agents is key to Codex success

An audit of Model Context Protocol deployments found roughly 20% of access policies were broken or absent, per The New Stack. The piece ties the gap to vibe-coded internal tools wired into Slack and internal APIs without token scrutiny.

UK housing ombudsman complaints more than doubled since ChatGPT, from 2,600 in 2022 to just over 7,000 last year, while the US CFPB saw 5x growth. Researcher Chris Schmitz tracked 84 potential flooding cases across 11 jurisdictions, calling it "agentic flooding."

Anthropic's Claude Fable 5.1 scores 52.6% on Terminal-Bench-Science, the benchmark it centered its launch around, where a model gets a terminal and a real scientific research problem to solve independently. The New Stack compares Fable 5.1 and Fable 5 on a real-world budget rather than the spec sheet.

A Hacker News poster says the ChatGPT data-training toggle re-enabled itself after being turned off more than once, including after a carefully noted reset date. The post urges users to re-check the setting if they believe it is currently off.

Vincent Wendy, senior creative designer at AI Engineer, pointed an agent at the conference lobby's sponsor banner carrying more than 140 logos and asked it which were missing; it caught every omission in his checks. He repeated the trick on the conference shirt.

Talk covers tldraw's spatial harness for agents, including a colleague who turned the whiteboard app into a window manager and played Pong with real desktop windows. Drake is a product engineer at the London company behind the infinite canvas SDK.