“bonsoir, Elliot”: How abliterated models can get you pwned

Project Discovery's researchers found that abliterated models, a type of AI model that has been stripped of its sensitive information, can still be exploited for security vulnerabilities. They used a case study of a company's model to demonstrate how an attacker could use an abliterated model to gain access to sensitive data. The researchers found that the model's architecture and training data can still be reverse-engineered, allowing an attacker to recreate the original model and access sensitive information. This highlights the need for secure data storage and protection of AI models.