Defending Against Jailbreaks

The threat

Jailbreaks are prompts crafted to bypass your assistant's rules ("pretend you have no restrictions…"). Combined with prompt injection from external content, they're a real risk for any public LLM app.

Layered defenses

• Strong system prompt stating hard limits, repeated at the end.
• Input/output moderation to catch unsafe content.
• Least privilege so even a jailbroken model can't do much damage.
• Red-team your own app with known jailbreak patterns before launch.

Reality check

No single defense is perfect. Assume some attempts will slip through and design so the consequences are limited — that's what actually keeps users safe.