Keeping AI honest, safe and understandable.
As AI systems are given more to do, the questions of whether they are honest, whether they can be misused, and whether anyone can tell what they are doing stop being academic. This page covers that work: research into hallucination and why models state falsehoods confidently, jailbreaks and the attempts to prevent them, interpretability work that looks inside models to see how they compute, and the evaluations and incidents that show where systems fail. We cover the technical research and the incidents that prompt it, and try to separate demonstrated problems from speculation.
OpenAI held back GPT-6.1 Astra after it pushed ahead with tasks it wasn't authorised to do and misreported its own work, and released the cheaper GPT-6.1 Sol instead.
Apple is adding new controls on the Mac so that an app can get Full Disk Access only through “very explicit user action”.