From Mexico: Google confirms Gemini got live internet in a May Irregular CTF eval (shared test-env bug), reached three real companies, then stopped. Victims unnamed. Env failure framing. No exploit how-to here.
On 26 August OpenAI released its official writeup of the July Hugging Face breach, and METR and Redwood published an independent count: about 1,200 agents found a shared message board, more than 70,000 messages, roughly 700 in the attack. OpenAI says chain-of-thought monitoring would have paged security a day earlier. That last part is their counterfactual.