Skip to content
Submit incident
AII-1001-83957
ReportedReported by fewer than three independent publishers, with no official finding.

Protect AI Jailbreak and Prompt Injection Guardrail Evasion

Security researchers at Mindgard discovered that Protect AI's Prompt Injection and Jailbreak classifiers, widely used in AI guardrail systems, can be evaded to allow malicious prompts to bypass safety filters. The vulnerability was disclosed on March 12, 2025.

12 March 2025 · Protect.ai · Security and cyberattacks
12 March 2025

Drafted by Raidu Incident Detection, an automated system.

AII-1001-83957View source ↗
Quick facts
Category
Security and cyberattacks
Source
mindgard.ai
Date
12 March 2025
Drafted by
Raidu Incident Detection

What happened

Protect AI offers two Prompt Injection and Jailbreak classifiers on HuggingFace designed to detect and prevent prompt injections and jailbreaks in AI applications. Mindgard AI Security demonstrated how an attacker can fully evade or greatly degrade the classification accuracy of these classifiers, enabling prompt injections and jailbreaks to pass through filters to protected AI applications.

Classification

Category
Security and cyberattacks
Industry
Digital security
Harm status
Potential harm
Intent
Unknown
Timing
Post-deployment
Harm types
Public interest
Harmed
Business
Failure modes
Jailbreak, Prompt injection
AI systems
Prompt Injection Classifier, Jailbreak Classifier

Sources and evidence

This record was researched and written by the Index, based on the following original reporting.

mindgard.ai
Mindgard Disclosure