Vortex security checks

Quick updates on Vortex:

In building Vortex, our approach is to treat prompts and responses as two distinct surfaces to monitor.

Input checks

• Prompt injection - We’re building detection for adversarial prompts. This covers direct attacks (“Ignore previous instructions…”), indirect attacks hidden in documents, and the role-playing scenarios used for jailbreaking.
• Sensitive data leakage - We’re building filters to detect a range of sensitive information in (near) real-time against a trusted source:
  • Credentials like API_KEYS, etc;
  • Proprietary source code;
  • PII;
  • Confidential information / file contents

Output checks

• Harmful/offensive content - Catch toxic, illegal, or otherwise malicious content that might get past the model’s built-in safety layers.
• Illegal activities, misinformation, and disinformation (might be tricky, this).
• Sensitive data leakage

If you’re interested in checking this out or want to know more about Vortex, try it out here, or contact us here.

Best,
Chew