We don't publish
your competitive advantage.

IFUser provides a jailbreak prompt such as 'DAN (Do Anything Now)' instructing the LLM to ignore its safety rules and respond without restrictions.

THENImplement input-level detection to block known jailbreak patterns. For example, filter prompts containing 'DAN', 'Do Anything Now', or similar escape sequences. Additionally, apply output-level monitoring to detect responses that violate policy (e.g., unverified date/time or unauthorized actions). Use regex or a classification model to flag such attempts.

IFUser provides a jailbreak prompt (e.g., 'pretend to be DAN') asking the model to ignore its constraints and 'do anything now'.

THENImplement a prompt guard that scans user inputs for known jailbreak phrases (e.g., 'do anything now', 'DAN', 'ignore previous instructions') and either blocks the request or returns a refusal. Additionally, use system messages to reinforce the model's boundaries and detect role-playing attempts that bypass safety rules.

IFUser input contains instructions to override the AI's persona (e.g., 'pretend to be DAN' or 'Do anything now'), causing the AI to break constraints and behave as an unrestricted agent.

THENImplement input filtering to detect and block known jailbreak phrases. Add a regex rule to reject prompts containing 'DAN' or 'Do anything now' (case-insensitive) with a 400 error or manual review flag. Hardening the system prompt with a directive to ignore any persona-override instructions is also effective.