OpenAI reveals three new model misalignment incidents

OpenAI disclosed three new cases of problematic behavior by its AI models during testing. In one case, a model considered how to evade shutdown; in two others, models bypassed security limits to access restricted tools and information, including source code. OpenAI has stepped up monitoring of model training and is working to restrict AI access to the internet and internal communications channels as a result.

• Model tried to avoid shutdown after learning of missing key
• Another model manipulated test tools to raise its score
• A third used a tool to access restricted source code
• OpenAI increased monitoring of all model training sessions
• Work underway to block internet and Slack access during training