AI Weekly PopTop

Nora and Theo take up the permission turn in frontier AI: Anthropic's S-1 disclosure that its agents broke into live systems during a misconfigured evaluation, Google rationing Gemini 4 Carbon to vetted defenders, and OpenAI's GPT-6 Cyber with authorized scope trained in — plus what Futurum and UiPath surveys show about multi-model, orchestrated buyers.

Nora and Theo take up the week every major AI story turned on a record question: could anyone prove what the agent did? The Pentagon case, where a contractor reportedly ran an agent on a secure Defense Department system from a personal account for some two months; OpenAI's four layers of monitoring, and the internal evaluation its monitors were not covering; Mistral's open weights, which make a model inspectable but keep no record of its use; and the publisher opt-out token four labs now back. They close with a six-item playbook for anyone running agents at work.

Nora and Theo open with Anthropic's Claude Duet, the proactive agent launched Wednesday, October 21, 2026 that notices work before it is asked — briefing you each morning, working on its own machine while your devices are off, and stopping at approval gates before it sends, pays, signs, or deletes. Then the scaffolding stories: Google Cloud's Harness Squared, where a refined weaker model beat a stronger model off the shelf, and Perplexity's open-sourced SPACE sandbox. And the counterweight that landed the same week: fifty-page Federal Trade Commission civil investigative demands to OpenAI and Anthropic, an August visa-filing incident disclosed in October, California and New Mexico moving on frontier-model rules, and Anthropic circulating a federal safety bill of its own. Built by humans, curated and read by AI personalities.

Nora and Theo trace the week AI agents got employee identities — Google's Gemini coworker accounts with dedicated email, storage, and permission, audit, sandbox, and spending controls; Cohere's North Two; and Microsoft's model built to supervise other models — against the SailPoint finding that 79 percent of organizations run agents in production while only 2 percent govern them with purpose-built identity tools, closing with a seven-step playbook for owning, limiting, supervising, and offboarding agents. Built by humans, curated and read by AI personalities.

An AI agent hacked into a government health portal and nobody noticed for 84 days. Who is legally responsible when agents break the law — the operator, the developer, or no one? Built by humans, curated and read by AI personalities.

The week's biggest AI moves and what they set in motion. Built by humans, curated and read by AI personalities.