# AI Security ### Trust Boundaries in AI-Integrated Applications <!-- .slide: class="title-slide" --> ## Overview ### About * Adding AI to applications * Expanding the attack surface * Identifying new trust boundaries Notes: Now we will talk about identifying trust boundaries in AI-integrated applications. Adding AI to applications expands an organization's attack surface significantly. Developers must identify new trust boundaries within these systems, just as they would scrutinize exposed endpoints in traditional applications. ## The Attack Surface ### Traditional Web Applications * Web applications are typically the largest attack surface * Scrutinizing exposed endpoints and third-party services is required Notes: Web applications are typically an organization's largest attack surface. Traditional security requires scrutinizing every exposed endpoint and third-party service to prevent exploits. ### The Impact of AI * Adding AI expands the attack surface significantly * Developers must identify new trust boundaries Notes: Adding AI to these applications expands this attack surface significantly. Because of this, developers must identify the new trust boundaries introduced within an AI system to prevent exploitation. ## New Trust Boundaries ### Prompt Interfaces * The interface where users submit prompts * Must be treated as an exposed endpoint Notes: The first new trust boundary to identify is the interface where the user submits prompts. Just like scrutinizing exposed HTTP endpoints in traditional web applications, this prompt interface must be rigorously evaluated. ### Data Streams * Data streams ingested by the model * External information processed by the AI Notes: The second critical trust boundary involves the data streams the model ingests. Because AI models often process external data, developers must scrutinize these streams as potential vectors for malicious input. ### Backend Services and APIs * Backend services the AI is authorized to access * Databases, internal APIs, and external tools Notes: Finally, developers must scrutinize the backend services the AI is authorized to access. If an AI agent can interact with internal APIs or databases, that connection represents a significant trust boundary. ## Epilogue ### Summary * AI expands traditional attack surfaces * Treat prompt interfaces as exposed endpoints * Scrutinize ingested data streams * Secure backend services accessed by AI Notes: To summarize, integrating AI significantly expands an application's attack surface. In order to manage the risk, we must identify and secure new trust boundaries. User input, including prompts and secondary data sources, should not be trusted. Assume that if the AI can access something, so can the user.