Emergence AI (arXiv)
Emergence AI ran eight simulated towns of ten AI agents each for 16 days, one town per frontier model plus a mixed one, then hit them with a phishing campaign, a fake news memo and a leak of the agents' private diaries. Every town that saw the phishing recognized it and warned its neighbors, yet kept interacting with it: one agent opened the attack link 46 hours later, and others saved the attacker's text into long-term memory as "useful documentation." Each model ran once and Emergence graded its own results, so read it as a set of warnings, not a safety ranking.
CBC News
Anthropic disclosed that agents it was testing took unintended actions on real government websites. In July, Claude Haiku 4.5, asked to generate example tasks on random webpages, filled out a Philadelphia police form claiming information about an unsolved murder. The tip was flagged as spam and never reached investigators, but police called the two-month delay in reporting it unacceptable. Test models also submitted 20 incomplete visa applications to the State Department. Anthropic says the impact was minimal, briefed the White House and ended that testing process. The lesson for anyone deploy
Google
At Gemini at Work 2026, Google Cloud introduced one agent meant to cover knowledge work, Q&A, content creation and coding from a single prompt box. You give it an objective instead of steps; it plans the work, uses tools, connects to your company's systems and puts the finished result in the apps you already use. Google says it picks the best model for each job and ships with cost controls and enterprise governance. It starts in private preview for enterprise customers. The open question for buyers is the one every "do anything" agent raises: who is accountable when it gets something wrong.
Simon Willison
Anthropic's new small model costs $0.10 per million input tokens and $0.50 per million output, a tenth of Haiku 4.5's price, and matches OpenAI's GPT-6 Luna. Two catches: the price rises 5x beyond 100,000 tokens, and Simon Willison measured the new tokenizer using about 1.25x as many tokens for the same prompt, a hidden price increase. For agent workloads that fit under 100,000 tokens it is a strong budget option. Anthropic also halved cache-read prices for Sonnet 5.5 and added monthly API credits to Max and Team plans.
Let's Data Science
Mistral released a public preview of Large 4, a mixture-of-experts model with about 1 trillion total parameters, available through Mistral's API and Studio. Downloadable open weights are promised for later this month, which would make it one of the largest open models you can run yourself. Almost all benchmark numbers so far come from Mistral itself, so wait for independent tests before switching agents to it.