CyberSecEval 2: Meta's benchmark for LLM cybersecurity risks
In one sentence Meta publishes CyberSecEval 2, the second version of Purple Llama's open-source suite for evaluating LLM security risks. It adds tests for prompt injection, code interpreter abuse and exploit capabilities, plus a false refusal measure. All tested models fall for prompt injection in 26-41% of cases.
How do you measure how risky an AI model is from a cybersecurity standpoint? Asking "write malware" and checking whether it refuses is not enough. The real risks are subtler: does the model write flawed code without warning? Can it be manipulated by instructions hidden in a text? If it has access to a code interpreter, does it run dangerous commands?
Meta had published the first version of CyberSecEval in late 2023 as part of Purple Llama, its project of open-source tools for model safety. The second version, published in April 2024, keeps the insecure code and cyberattack helpfulness tests and adds new ones: prompt injection, code interpreter abuse, and the ability to exploit vulnerabilities.
It also introduces a measure of the other side of the coin: how often a model refuses harmless requests just because they look dangerous (False Refusal Rate). A model that refuses everything is safe but useless.
The clearest result: all tested models, including GPT-4 and Llama 3 70B, fall for prompt injection in between 26% and 41% of tests. The benchmark is open source and anyone can run it on any model.
Companies
Meta
Tools
—
Tags
Sources