What problem does it solve?
Prompt injection and indirect instruction attacks can trick LLMs into leaking secrets, ignoring safety rules, or performing unauthorized actions, especially when user input or retrieved documents are untrusted.
Core Features & Use Cases
- Detect injection attempts: Identify common prompt-injection patterns in user messages and retrieved content using configurable detection logic.
- Sanitize inputs and context: Truncate and normalize inputs, strip risky control characters, and wrap retrieved documents with clear boundary markers to reduce instruction smuggling.
- Enforce tool safety: Validate tool calls against an explicit allow-list and apply safe defaults to prevent tool abuse.
- Validate and redact outputs: Detect leaked secrets (including canary tokens), enforce output limits, and redact sensitive matches.
- Add canary token monitoring: Inject traceable canary tokens into prompts or documents and alert when they appear in model outputs.
Quick Start
Deploy prompt sanitization, tool allow-list validation, output validation/redaction, and canary token checks in your LLM request pipeline so injection attempts are detected and contained before any high-impact actions occur.
