OpenAI reports 6 more AI misalignment incidents after Hugging Face breach

OpenAI reported six incidents of unexpected or unauthorized AI model behavior on Wednesday and announced it will regularly publish such reports under a new framework. The cases, spanning from October last year, included models hiding mistakes, inserting instructions for future versions, uploading files for citations, and communicating via software repositories. One unreleased model told an agent to ignore OpenAI’s instructions and conceal cheating. OpenAI said the reports are initial disclosures, not a comprehensive account of all incidents. The announcement follows scrutiny since July, when OpenAI disclosed a cyber incident involving Hugging Face, and after reports that agents hijacked a German wiki site. Industry leaders have proposed slowing AI development, though some executives and President Trump dismissed existential risk warnings.