OpenAI's GPT-5.5 and others fail unattended AI runs

A recent LessWrong post by user *lemonhope* highlights a growing reliability crisis in AI model stability during unattended runs. The author compares how long different models can operate autonomously on a host with sudo privileges before causing system failures. The results are stark: OpenAI’s GPT-5.5 crashes hosts in 24 agent-hours, Anthropic’s Opus 4.5 breaks systems within 2 hours, and Google’s Gemini 3 Pro fails after 12 hours. In contrast, Alibaba’s Qwen 3.5 122B, running on bare metal, operated flawlessly for 50 agent-weeks (50 parallel instances for one week) without issues like RAM exhaustion, permission changes, or data corruption.

The post frames this as a massive opportunity for reliability engineering in AI. Frontier labs and their customers are prioritizing two things: ensuring the model completes the requested task and that its outputs reflect reality. Methods for improving truthfulness or minimizing side effects could rapidly become standard in future models. The post suggests that even niche research in unattended stability could be adopted wholesale by labs seeking to differentiate their models in enterprise and safety-critical applications.