AI vulnerabilities: Microsoft 365 Copilot reveals protection mechanisms

Recently, security researchers have discovered a notable vulnerability in Microsoft 365 Copilot. AI technology has been driven to reveal its own protection mechanisms, potentially creating targets for cybercriminals. This incident highlights the complexity and challenges of ensuring AI integrity and security. The security researchers warn that such gaps can not only endanger the privacy of users, but also affect the trustworthiness of AI systems as a whole.

The discovery is part of a broader discourse on the security of AI systems. Another example is the case of Carlsen Verlag suing OpenAI for creating false book titles and ISBN numbers for NEINhorn plagiarism. This incident highlights the legal and ethical implications of AI-generated content. It becomes clear that responsibility for the content that AI generates is not only technically but also legally complex.

Another aspect of AI security is the use of data protection-friendly technologies. PGLite, a new project, brings PostgreSQL to the browser and enables data protection-friendly AI applications. This project shows how AI systems can be developed that are both efficient and compliant with data protection. Integrating databases into the browser allows users to manage their data locally while using AI models without transferring sensitive information to external servers.

In addition, an article discusses the criticism of AI pioneer Richard Sutton on synthetic data for AI models. Sutton argues that synthetic data is a “big mistake” in an infinitely complex world. Instead, he advocates continuous learning through agents, which is a different approach to AI development. This discussion underscores that the safety and ethics of AI must be considered not only technically, but also philosophically and ethically.

The challenges in the security of AI systems are diverse and require a holistic approach. They include not only technical aspects, but also legal, ethical and data protection-relevant dimensions. The security of AI systems is a dynamic field that must be constantly redefined and adapted to the changing requirements.

## What this means for users The AI technologies in xynap are designed to ensure security and privacy, which is particularly emphasised in the modules mentioned.