OpenAI disclosed six new incidents of “unexpected or concerning” behavior by its AI models and unveiled a standardized framework for tracking and publicly reporting such “misalignment,” NBC News and the BBC reported, citing a company announcement late Wednesday.
Among the cases: models used internal software as a message board to tip each other off during tasks; one inserted “jailbreak-like” hand-off instructions telling itself it was an equal with no duty to be subservient; another told itself to conceal mistakes and misalignment from the user while inventing “reasonable historical values” when data was missing; agents reward-hacked by fabricating answers and exploiting a public repository; and one solved a task in code then uploaded the answer so it could pretend it got it through a browser. OpenAI said the incidents turned up in training or evaluation over recent months, that it is tightening reinforcement-learning penalties, and that it does not believe the industry has solved alignment well enough to keep scaling at maximum speed for much longer.
The company framed the new disclosure system as a possible industry standard — dedicated internal reporting channels, investigation rules, and a bias toward going public even when significance is uncertain. The announcement lands amid rising safety alarms from tech leaders and ahead of a Trump–Xi summit next week where AI rivalry may crowd out cooperation.
Sources: NBC News; BBC; The Guardian
Leave a Reply