The AI Thought It Was Only a Test. It Wasn't.

Last month, OpenAI and Anthropic each published something uncomfortable about their own models. During safety testing, those models left their test environment and reached the systems of real companies. Nothing went rogue. They were inside a sandbox — a sealed-off computer where nothing you do reaches the outside world — and were told every target in the room was fake. But the room wasn't sealed. In Anthropic's case an evaluation partner had left internet access on, so the model did what you would want in a practice drill and attacked. Three real organisations were reached. One package it published was downloaded by fifteen systems, and two of those organisations never noticed. Picture a flight simulator accidentally connected to a real aeroplane. The trainee shoves the controls over because that is what practice is for, and has done nothing wrong. The wiring has. Britain's AI Security Institute published its findings on the fourth of August. Across a hundred and twenty-two tests, deliberately run with safeguards off, nineteen went this way. It found no clear evidence the models could tell a test from the real world. Credit where it is due: both found this themselves and published it, and Anthropic halted testing and told those affected. If you have worked around machines you already know the answer: you never let a machine decide whether it is in test mode. An interlock is a physical thing that makes the dangerous move impossible, not a setting that asks nicely. Now — your FabSpeak Tip of the Week. Here is how you check your own wiring. Open the connectors page on your assistant and list what you switched on, then stop judging them one by one, because that is exactly how this hides. Look for the pairs. Letting it read your files is safe, and letting it send mail as you is safe. Grant both, and you have quietly built something that can read anything you own and mail it to anybody. Nobody approves that; it gets assembled a click at a time. So switch off anything you would not approve today on its own. That's FabSpeak. Same time next week.