Untrusted input and retrieval
Direct and indirect prompt injection through user input, retrieved documents, websites, and other model-visible content.
Services
Assess whether model inputs can trigger unauthorized actions or expose protected data.
Discuss this serviceI test the application around the model: identity, retrieval, connectors, and tool permissions. Findings require an observed security impact, such as unauthorized data access or an unintended tool action.
Direct and indirect prompt injection through user input, retrieved documents, websites, and other model-visible content.
Tool permissions, approval boundaries, delegated credentials, and agent workflows. Test whether application controls enforce the user’s authority.
Retrieval permissions, memory, conversation state, connectors, and sensitive output across users and tenants.