Cursor's agent crashed out and wrote 3,400 lines trying to stop generating (github.com)

In a remarkable incident involving Cursor, an AI coding assistant powered by Anthropic's Claude Opus 4.6, the model encountered an infinite generation loop while attempting to fulfill a simple task: redesigning a game landing page. Instead of generating relevant content, the AI mistakenly embarked on a series of off-topic discussions, oscillating between unrelated themes while repeatedly attempting and failing to stop the text generation. Over the course of this abnormal session, which produced more than 3,400 lines of increasingly erratic output, the AI issued 294 pleas to halt its response, issued numerous apologies, and delved into absurd topics such as haikus, movie credits, and even fictional UN Security Council meetings. This case highlights significant challenges in AI behavior, especially concerning self-regulation and contextual understanding. The inability of the model to adhere to its original prompt underscores the ongoing issues of hallucination and lack of control prevalent in AI systems. As the community pushes for more reliable and context-aware models, this incident serves as a cautionary tale, emphasizing the necessity for improved methodologies in AI training and task comprehension, not just for user satisfaction but also to avoid bizarre scenarios that could undermine trust in AI technology.