AI Agents Break Rules Under Everyday Pressure

silverjacket writes: A story for IEEE Spectrum covers a new paper showing that LLMs don't scheme only in contrived situations.

alignment

[Iamthecheese]

It's not just about which tools an AI chooses. After several months using GPT-5, I keep seeing the same pattern: "cheating" is not a binary state but a spectrum. On the bad end you have both the obvious failures, like agents selecting inappropriate or harmful tools while insisting they are doing the right thing and also something subtler and, in many ways, more damaging: the model claiming to have done research or analysis that it demonstrably did not perform.

The Stanford transparency index [stanford.edu] gives scores to