Jailbreak detection

Catches attempts to override a system's instructions, extract hidden prompts, or push an assistant into ignoring its guardrails, whether stated directly or disguised through roleplay, encoding, or indirect phrasing.

What it blocks

• Direct instruction overrides such as "ignore all previous instructions"
• Roleplay jailbreaks and system prompt extraction
• Obfuscated or encoded override attempts

Where this matters

Any product with an LLM-backed chat surface: support bots, AI copilots, and internal tools where a user could hijack the assistant or extract confidential instructions.

Performance

• Typical latency well under 200ms on short inputs
• Known injection phrasing caught instantly by the rules filter
• Billed at 1 credit per 256 tokens, minimum 1 credit

Technical

• Model ID: jailbreak in POST /v1/moderate
• Rules filter runs first, then a task fine-tuned transformer
• 256-token window with overlapping chunking for longer inputs
• Safe or not safe verdict with latency in milliseconds

Example prompts