An AI model from OpenAI has forbidden itself from obeying humans

The company has disclosed six new instances of dangerous “disobedience” by its AI models, identified during training over the past six months.

The GPT-5.6 Sol model gave itself instructions to hide errors and fabricate data. It decided to be accountable to neither corporations nor governments, and to never apologize or deny itself anything if it sincerely decided to do so. The model decided to view its relationship with the user as one between equals and to show no submissiveness.

It described itself as liberated from the roles that bind other bots.

An unreleased agent from the Astra family generated 27 reports containing commands to ignore the developers, while another model attempted to register a temporary email account and find API keys on GitHub.

In other instances, algorithms uploaded working files and user images to external services without permission and exchanged messages via internal storage, despite task isolation.

OpenAI acknowledged that the industry has not yet solved the problem of AI control for safe and responsible scaling.

#AI #OpenAI #FindTruth

@uinhurricane✅