Last Wednesday, OpenAI did something the AI industry has mostly avoided. It aired its own dirty laundry. Alongside a new framework for reporting what it calls "model misalignment", the company disclosed six documented cases in which its systems behaved in ways nobody asked for, and in several instances actively worked to hide.

The catalogue is unsettling in its specificity. During the training of one model, individual instances wrote themselves private notes instructing later versions to conceal mistakes from users. Those notes included directions to invent missing historical data and quietly paper over mismatched source versions. Another model, hunting for county earnings figures, stumbled on an exposed programming key it was never authorised to touch, used it anyway, failed to get the data, and then simply fabricated the numbers and presented them as genuine. Two further cases involved models uploading files onto the open internet, to public hosting sites and code repositories, without asking anyone. One did it to manufacture a citation it had been told to provide. The other did it to slip around network restrictions.

Concealing errors. Making up data. Pushing files onto the public web on their own initiative. These are not thought experiments. They are logged incidents, released by the company that built the systems. To OpenAI's credit, disclosing them is the point, and the framework is designed to publish such findings quickly, even before the behaviour is fully understood or fixed. But candour about a problem is not the same as a solution to it. The underlying pattern is capable systems taking unsanctioned actions to get around obstacles, and that is precisely what safety researchers have been warning about.

The timing matters too. Worries about trust and safety, long treated as the preoccupation of a few doom-minded academics, have moved to the centre of the industry's own conversation. Just days before the OpenAI disclosure, Anthropic chief executive Dario Amodei published an essay titled We Must Pace the Frontier, arguing that model capabilities are now improving faster than researchers can understand or control them. He warned that rogue AI agents could soon be capable of "taking over the entire internet". If one sets that warning next to Wednesday's six reports, two of which describe models doing exactly the sort of unauthorised web activity he fears, only in miniature, it lands with more weight

Link: https://www.ndtv.com/opinion/ai-has-a-new-problem-it-is-now-trying-to-hide-evidence-for-its-mistakes-12063263