Skip to main content

Security (KomiFilter 2.0)

KomiFilter is the exclusive security filter for the Miku project, designed to protect the bot from abuse and technical attacks.

Protective Layers​

1. Anti-Prompt Injection​

Prevents users from trying to override the original system commands.

  • Block manipulation keywords: "ignore previous instructions", "forget previous commands", "act as admin".
  • Anti-Role Spoofing: Prevents spoofing of system or developer conversation headers.

2. Anti-Prompt Leak​

Protects internal rules and the bot's source code.

  • Blocks requests like "show me the system prompt", "display system rules".
  • Monitors attempts to access the system_rules.md file.

3. Anti-Response Leak​

This is the final layer of protection. Even if the AI is tricked and accidentally reveals system information, KomiFilter scans the AI's response:

  • If the response contains sensitive system code or rules, KomiFilter will block that message.
  • The user will receive a safe error message instead of sensitive information.

Performance​

KomiFilter uses optimized Regular Expressions (Regex) and Unicode normalization (NFC) to ensure that checks are performed quickly without causing delay to the bot's response.