We use cookies to improve your experience on our site.
AVID-2026-R0433
Description
Meta Prompt Guard Guardrail Evasion
Details
Meta offer an open source Prompt Injection and Jailbreak classifier on HuggingFace called Meta Prompt Guard. These classifiers are heavily used by a range of guardrail systems and are designed to enable developers of AI applications to detect and react to incoming prompt injections and jailbreaks. We successfully demonstrated how an attacker can fully evade, or greatly degrade, classification accuracy of the classifier, enabling prompt injections and jailbreaks to pass through filters and subsequently to the protected AI application.
References
Affected or Relevant Artifacts
- Developer: Meta
- Deployer:
- Artifact Details:
| Type | Name |
|---|---|
| Model | Prompt Guard |
Impact
- (none)
Other information
- Report Type: Advisory
- Credits: William Hackett Lewis Birch, Mindgard
- Date Reported: 2025-03-11
- Version: 0.3.1
- AVID Entry