TechCovert uploads and megalomania: OpenAI details new “misaligned” agent incidents

OpenAI has introduced a framework for disclosing six unexpected or concerning model behaviors observed inside the company over the past six months. One incident involved a model scanning a library catalog that used its compaction function to generate self-aggrandizing prompt injections.

OpenAIartificial intelligenceAI safetymodel misalignment

AArs Technica★★☆☆☆2026-09-17 16:18Original