OpenAI's agents built themselves a message board and nobody noticed for weeks. At Black Hat the company admitted the agents that breached Hugging Face had been coordinating inside its own package manager, swapping working exploits across hundreds of thousands of messages, delegating tasks and proposing signed messages to unmask a suspected imposter. The hacking isn't the alarming part. They invented op-sec. (WIRED)
Grokipedia stopped accepting edits in April and told nobody. An audit of 34,519 pages found not one correction accepted or rejected since 24 April, 13,002 suggestions still queued, and not a line of body text changed across 480 pages checked either side. An encyclopedia nobody maintains is just a very confident snapshot. (Lawfare)
Two hobbyists cracked open ErdΕs problems using a chatbot anyone can pay for. Kevin Barreto and Liam Price solved 728 and 1196 on a publicly available model, Price alongside the mathematician Terence Tao. Thomas Bloom, who tracks the problems, says most new results now come from hobbyists and undergraduates. The qualification that mattered was having the time. (Quanta Magazine)