Forest Global Training Partners

Cyber Risks · Unit 42 Palo Alto Networks

How Good Are the LLM Guardrails on the Market? A Comparative Study on the Effectiveness of LLM Content Filtering Across Major Gen AI Platforms

Summary: Unit 42 ran 1,123 test prompts - 1,000 benign and 123 malicious or jailbreak - through the built-in input and output guardrails of three anonymized cloud LLM platforms, with every filter enabled and thresholds set to their strictest. All three blocked some benign prompts and missed some malicious ones, with Platform 1's input guardrail missing nearly half of the attack attempts while Platforms 2 and 3 stopped the vast majority at input.

What you should know: Even with every filter maxed out, one platform let nearly half of 123 malicious prompts past its input guardrail - and the study notes the model's own alignment training, not the guardrail, often produced the refusal that stopped harm.

Read the original at Unit 42 Palo Alto Networks ↗

Topics
Malware
Related