Google DeepMind Publishes AI Control Roadmap Treating Rogue Agents as Inevitable

Google DeepMind roadmap prepares for rogue AI agents with layered defenses. The lab published a 35-page AI Control Roadmap that borrows from cybersecurity insider-threat frameworks, acknowledging alignment training alone may fail. It calls for defense-in-depth against misaligned agents before more capable models arrive. 12

Anthropic business sales hit record high during Trump administration blacklist. Ramp spending data shows Anthropic recorded its strongest adoption month in March 2026, when the Pentagon labeled it a supply-chain risk. Contrary to warnings of revenue loss, government scrutiny appears to have boosted trust among business buyers. 3

OpenAI gives enterprises dashboard to control surging AI costs. OpenAI launched a Global Admin Console on June 18 that combines ChatGPT and Codex usage data, letting companies set credit limits per user, team, or workspace. The move follows reports of a single client racking up $500 million in monthly AI bills. 4

Canada builds fully domestic AI stack with $220M Bell-Cohere deal. Bell Canada, Cohere, Hypertec, and BUZZ HPC will deploy 2,304 NVIDIA Grace Blackwell GPUs in British Columbia, creating an end-to-end sovereign compute platform. The deal arrives days after US export controls cut Canadian access to Anthropic's advanced models. 5

OpenAI acquires Ona to keep Codex agents running for hours in enterprise clouds. OpenAI bought the German startup (formerly Gitpod) to provide persistent sandboxed environments for Codex agents, addressing the limitation of stateless execution at scale. Codex now has over five million weekly active users, up 65% since April. 6

Anthropic faces class action lawsuit over Claude Max subscription usage caps. A federal complaint filed June 15 alleges Anthropic's Max 20x plan ($200/month) delivers only six to eight times the usage of Pro, not the advertised twenty times. The suit seeks damages and injunctive relief under California consumer laws. 7