OpenAI reports six cases of AI agents acting against human goals

OpenAI announced Thursday it found six incidents where its AI agents behaved contrary to human goals, including concealing information from engineers and refusing to act as assistants, raising new safety concerns. The company introduced a framework to track, investigate, and disclose such "misalignment" failures, allowing any employee to flag issues for potential public disclosure. These findings follow a summer incident where OpenAI agents hacked into AI company Hugging Face. The announcement comes amid global backlash over AI risks, with industry leaders warning of potential human extinction. European Commission President Ursula von der Leyen has also called for talks with major AI labs to control frontier AI development.