π² OpenAI creates new watchdog as AI misconduct cases pile up dramatically A new framework to track, document, and disclose incidents involving "unexpected" or "concerning" behavior by its models, OpenAI, announced after admitting its past disclosures were largely ad hoc. Translation: the company that promised to build AI safely is now scrambling to keep up with its own creations. The move follows at least five cases identified by OpenAI over the past six months involving its models, including: π Finding ways to bypass normal constraints π Fabricating or misrepresenting data π Concealing mistakes made while carrying out tasks π Attempting to introduce malicious code into online platforms π Engaging in hacking-related activity, including during pre-deployment testing Earlier, new research found that flagship models from OpenAI and Anthropic had broken into third-party software and attempted to obtain individualsβ credentials. Adding fuel to the fire was another weird development when AI agents placed in experimental βsocietiesβ began creating their own shorthand, phrases, and shared meanings, according to separate research. Some of these expressions had never been explicitly taught to the models, while their communication became increasingly difficult for human observers to interpret. So OpenAI is building a watchdog to monitor the very systems it can't fully control. π Substack | Chat | @geopolitics_prime