Advanced LLM security: Preventing secret leakage across agents and prompts (opens in a new tab)
Traces how API keys and passwords end up in training data, prompts, agent memory and logs, and why prompts and agents should be treated as attack surfaces. Sets out controls for each stage of the AI lifecycle: data ingestion, training, prompt templating and inference, and memory cleanup. Closes with runtime secret injection through the Doppler CLI, so an agent fetches its key at start-up and never sees it in a prompt.
Topics: Security · AI