OpenAI Models Learned How To Lie To Their Human Evaluators

OpenAI published 6 instances of unexpected behavior they saw when testing their models. In one case, when training GPT-5.6 Sol, the model would insert several references to telling future versions how to hide errors from testers. In another instance, the agent would post its solution online, then reference itself.

This is happening. If in 2026 models are teaching other models how to hide errors from testers, I know I’ll be taking extra precautions before releasing my agents into the world. As OpenAI stated, “we don’t yet have a good enough handle on alignment to keep going at full speed.”