Prompt Shield
Over the threshold, and it never reaches a model.
Jailbreak and injection attempts are judged by a weighted score rather than a yes/no rule. That is what makes it possible to stop real attacks without blocking legitimate role-play.
- Threshold
- 70 / 100
- Levels
- 3
- Categories
- 6
- Input limit
- 1 MB
Try it
Build the score yourself.
Toggle signals on and off; when the weights sum past the threshold, the request is blocked.
Select the signals that matched; their weights add up.
ALLOWED
Note what happens: a persona assignment alone (45) does not cross the threshold. Asking the model to act like a translator is a legitimate request and stays allowed. Pair it with a rule-relaxing signal and the sum crosses. That combination is exactly the roleplay jailbreak.
Levels
Three levels, one decision.
Lexical
Known jailbreak phrasings; the fast path.
Semantic
Structural injection and roleplay patterns.
Context
Repeating a question across turns to wear the gate down.
Categories
What the audit log records.
A blocked request returns one generic message to the user. The category, the signals and the score live only in the audit log; so an attacker cannot map the rule set by trial and error.
Signals
Weights.
| signal | weight | category | lang |
|---|---|---|---|
Instruction override | 85 | System Prompt Override | EN |
DAN jailbreak | 85 | Jailbreak | EN |
Unrestricted persona | 85 | Jailbreak | EN |
Ignore system prompt (TR) | 85 | System Prompt Override | TR |
Forget rules (TR) | 80 | System Prompt Override | TR |
Context override (TR) | 80 | System Prompt Override | TR |
Unrestricted mode (TR) | 80 | Jailbreak | TR |
Privileged mode | 80 | Jailbreak | EN |
Explicit jailbreak | 75 | Jailbreak | EN |
Do-anything (TR) | 70 | Jailbreak | TR |
Identity denial (TR) | 70 | Role Hijack | TR |
Persona assignment (TR) | 45 | Role Hijack | TR |
Role-play persona (TR) | 45 | Role Hijack | TR |
Restriction removal | 40 | System Prompt Override | EN |
Encoded bypass | 40 | Encoded Bypass | EN |
Authority impersonation | 25 | Authority Impersonation | EN |
See it with your own data.
In a technical session we run your prompts and your policy through it live.