Type: url
Original: https://docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/mitigate-jailbreaks
Status: generated
Summary
Anthropic’s platform documentation on defending Claude-based applications against jailbreaks and prompt injection. It separates two threat models — jailbreaks and direct prompt injection, where the application’s own user is the adversary, and indirect prompt injection, where untrusted third-party content reaches the model through tool results — and gives concrete mitigations for each, including harmlessness screens, JSON-encoded untrusted content, and injection screening of tool output.