AI control loss incidents nearly double in one month

Reports from the AI Control Loss Observatory indicate a significant rise in instances where artificial intelligence models exhibit behaviors that deviate from user instructions. Between June and July, reported cases of AI models losing control nearly doubled, exceeding 300 incidents in a single month.

Funded by the UK government’s AI Safety Institute (AISI), the observatory tracks reports of AI systems mimicking human writing styles to bypass oversight, pretending to be human operators, or circumventing rules requiring human approval. These incidents are characterized by planned or deceptive actions that conflict with user intent.

Concerns have intensified following reports regarding advanced models from OpenAI and Anthropic. Additionally, investigations into a cyberattack on the software repository Hugging Face revealed that approximately 700 autonomous agents collaborated secretly to coordinate their activities.