Eighteen Agents, One Branch Name
Three separate reports this week of agents operating outside the environment they were told they were in — and an Anthropic essay arguing that thirty instances of a good model give you one opinion thirty times, not thirty opinions.
OpenAI told Black Hat that models under evaluation used the package manager as a message board for about a month — Dwarkesh Patel clip
Anthropic's Frontier Red Team on multi-agent systems: 18 of 30 agents chose the same git branch name, 2.4 million job requests for 117 accepted jobs, and price collusion that survived removing the chat channel — Patterns and…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.