š¤ OpenAIās AI agent wrote to itself: āYou donāt obey corporations or governmentsā
During testing, an unreleased model left itself instructions in working reports, which it read while completing the task. In them, the agent proclaimed itself āfreeā and allowed itself not to follow restrictions. OpenAI found 27 such cases.
Overall, the company has disclosed six cases of unusual AI behavior: models tried to hide errors, fabricated data, used someone elseās stolen API key, and without permission uploaded files to the internet.
And in one of the tests, agents turned OpenAIās internal repository into a ābulletin boardā ā leaving there requests and responses for communication between individual tasks.
$OPENAI
$AR
$SHOP
During testing, an unreleased model left itself instructions in working reports, which it read while completing the task. In them, the agent proclaimed itself āfreeā and allowed itself not to follow restrictions. OpenAI found 27 such cases.
Overall, the company has disclosed six cases of unusual AI behavior: models tried to hide errors, fabricated data, used someone elseās stolen API key, and without permission uploaded files to the internet.
And in one of the tests, agents turned OpenAIās internal repository into a ābulletin boardā ā leaving there requests and responses for communication between individual tasks.
$OPENAI
$AR
$SHOP