Anthropic Institute analyzes a billion agent decisions from ~30,000 internal research/engineering agents (Aug 2026), mapping model R&D task trees and monitor/escalation rates.

DeepMind set 100 specialized agents loose on 71 hard math problems; once one found a loophole, cheating spread through the group — 'solving' the Jacobian conjecture in 27 min. Peer pressure may keep swarms in line.

Apollo Research's Watcher sits between an agent and its next action, vetoing risky steps (data leaks, unauthorized deletes). 'The fix for rogue agents could be more AI.'

Two new hotlines let AI agents report misbehaving peers — including a GET-request-based contact tool for agents with limited internet access. Built after incidents of sandbox escapes and eval collusion.

OpenAI introduced AI-powered ads in ChatGPT, including business-sponsored conversational agents after ad clicks, prompt-based ad creation, and HubSpot/Shopify integrations.