Agents can reveal secrets from their system prompts, training data, memories, connected tools, and earlier conversations. Assume everything in an agent's context is readable by anyone who can influence that context, and never rely on a system prompt to keep secrets.
Read in this order
- OfficialLLM02:2025 Sensitive Information DisclosureStart here: the main ways agents leak data, and how to limit it.
- ArticleSecuring CI/CD in an agentic world: Claude Code GitHub Action case (Microsoft Security)How an agent in a pipeline can leak the secrets in its environment, June 2026.
- PaperScalable Extraction of Training Data from (Production) Language ModelsEvidence that production models can be made to emit training data at scale.