Pranjal Timsina | AI & Security

AI security must be considered from multiple angles - preventing users from exploiting your AI platform, making sure your AI does not produce harmful content, and keeping personal information of end users safe. "Non-deterministic automation" is taking over almost everything and new threat surfaces are emerging faster than we can keep up. The creation of environments to facilitate end users' AI agents taking actions on their behalf is also quite interesting from a security and liabilities perspective. Tool call poisoning, retrieval of "private" info that leaked into the training data, and training data poisoning are also topics I've planned to speak on. My presentation will mostly revolve around these ideas.