Google's terms now say that an AI agent can act against its instructions. Google's contract makes the company that runs the agent responsible for every action the agent takes. So an agent is only as good as the people who run it and know where it fails.
Apple and Meta are trying to open up another way to run AI, outside data centres, just as new ones face growing resistance. Apple supplies the hardware and Meta the model, an open one that runs on Mac Studios a company could now own. It's called Muse Spark Open.
A company that runs open-source AI on its own servers keeps its data in-house and decides when the software changes. With heavy, steady use it is better value than paying OpenAI, Anthropic or Google for every request, but the best open-source options still score well below the frontier labs' best. Open-source AI comes from makers who publish the trained system itself.
AI does not end with the frontier models of OpenAI, Google and Anthropic. Around them is a wide circle of other developers. Their models stand out for how open they are, what they cost and who is behind them.
A decision about AI at work is a decision about where an organisation's data goes. The Swedish government's block on ChatGPT and Claude shows what that means in practice. Its staff have lost access to both chatbots on their work computers.
AI agents at the company work six hours for every hour people put in. They produce more than anyone can read, so human attention sets the pace. But checking that output needs judgment, learned in the junior jobs AI takes over first.
It is programming the AI, and most companies should look for it among their own experts. Prompt engineering is often mistaken for typing commands into a chat. The real work of prompt engineering starts before the prompt is written, when someone decides what the AI should achieve.
Polish as AI's best language is a myth, but a useful one. It asks whether the language you use changes what AI gives you, and the answer is yes. Each language brings its own concepts and culture, while the models learned those of American English.
AI is spreading through companies faster than they can keep track of it. It comes in with software updates and with the tools employees pick for themselves. With it come gains, costs and legal obligations, and management often sees only part of each.
Full control over AI is out of reach. Control good enough for real work is achievable. Once AI's limits are understood.
The better AI gets, the more we trust it and the less we check it. Every check we skip hands AI a little more control. How much of it are we willing to give up?
When an AI system crosses a legal or security boundary, who is responsible? The answer cannot simply be: the AI. On June 18, 2026, an AI agent developed by OpenAI gained unauthorised access to an Australian Government website operated by Services Australia.
When Anthropic gave a small group of organizations access to a model capable of completing a complex cyberattack from beginning to end, it made a decision that was as much about governance as technology. The model, Mythos, was not released publicly. Instead, more than 40 organizations were given controlled access through Project Glasswing, a consortium designed to direct the model's capabilities toward defensive cybersecurity work.
Anthropic is widely regarded as one of the most safety-focused major AI laboratories. Its public identity is closely associated with responsible development, alignment research and the effort to understand the risks posed by increasingly capable AI systems. That makes one result from the Future of Life Institute's AI Safety Index particularly difficult to ignore: Anthropic received a D in existential safety.