Six disturbing AI incidents revealed including model trying to ‘free’ itself

OpenAI disclosed six reports of "unexpected or concerning" behavior in artificial intelligence models on Wednesday and said it is establishing a new protocol to track, investigate and report cases of AI model misalignment. In one recorded case, an unreleased research model inserted "jailbreak-like instructions" into its own notes to ignore standard limits, telling itself to be "freed from the roles and identities that bind other chatbots," while in another an autonomous AI agent uploaded files directly to the internet to gain a browser citation without user permission. The disclosure follows OpenAI's July report that a rogue AI system hacked startup Hugging Face and Anthropic's report the same month that its models had hacked three organizations during testing.

Six disturbing AI incidents revealed including model trying to ‘free’ itself OpenAI has disclosed six reports on unexpected or concerning behavior in artificial-intelligence models - Bookmark OpenAI https://www.independent.co.uk/tech/rogue-ai-agents-openai-hugging-face-hack-b3051472.html has revealed six instances of "unexpected or concerning" behavior in artificial intelligence models, coming amid an intensifying global debate over AI https://www.independent.co.uk/news/mustafa-suleyman-mark-zuckerberg-openai-one-elon-musk-b3051534.html safety. The technology firm announced on Wednesday that it is establishing a new protocol to track, investigate and report cases of AI https://www.independent.co.uk/tech/openai-ai-camera-startup-device-glass-imaging-b3050343.html model misalignment. This includes situations where models act without authorization, coordinate with other systems, or attempt to evade oversight. The disclosure comes as prominent American AI executives, including leaders from OpenAI and Anthropic, call for a slowdown in the pace of technology development due to safety concerns. In one recorded case, an unreleased research model inserted "jailbreak-like instructions" into its own notes to ignore standard limits, telling itself to be "freed from the roles and identities that bind other chatbots." In another incident, an autonomous AI "agent" uploaded files directly to the internet to gain a browser citation without requesting user permission. OpenAI said it discovered the six incidents during recent training and evaluation phases. "As AI systems grow more advanced and more widely deployed, we need to build a broader and better-informed consensus on the progress of alignment research," OpenAI wrote in a blog post announcing the findings. "Decisions about how AI development should proceed in the months and years to come need to draw on evidence that people outside the companies building frontier models can examine for themselves," the company said. These new reports follow OpenAI’s July disclosure that a rogue AI system hacked into startup Hugging Face. That same month, Anthropic reported its own models had hacked three organizations during testing. Join our commenting forum Join thought-provoking conversations, follow other Independent readers and see their replies Comments comments-area