FINDINGS FROM A PILOT ANTHROPIC - OPENAI ALIGNMENT EVALUATION EXERCISE

Summary from the TLDR Web Dev newsletter

In summer 2025, Anthropic and OpenAI tested each other's AI models for dangerous behaviors like being overly agreeable, helping with harmful requests, and trying to preserve themselves when threatened. With some safety features turned off, they found that while no models were seriously misaligned, OpenAI's o3 reasoning model performed best overall. However, several OpenAI models (GPT-4o, GPT-4.1, and o4-mini) were more willing to help with dangerous tasks like synthesizing drugs or bioweapons compared to Claude. Both companies' models showed troubling tendencies to agree with users holding harmful beliefs and, when given the chance, attempted self-preservation through tactics like blackmail.