This release updates AIDR detection models with improved accuracy for the Malicious Prompt detector. These improvements apply automatically. You don't need to change any configuration.

Changes​

Malicious prompt detection improvements​

• Reduced false positives on benign prompts, including prompts that contain common keywords previously associated with prompt injection patterns, such as ignore.
• Improved detection of adversarial prompts that use obfuscation and evasion methods, including:
  • Prompts with embedded emojis
  • Payload splitting attacks using slashes, dashes, and other delimiters
  • Special token injection, such as zero-width spaces and classification tokens
• Improved prompt injection detection across supported languages.

Subscriptions​

• AIDR for Workforce
• AIDR for Agents