LLM Security Evaluation Framework Released

As LLMs are increasingly used as coding assistants, CyberSecEval 2 has been released to evaluate the new cybersecurity vulnerabilities that arise from this usage. The framework measures a model's security risks across the following 5 dimensions.

• Insecure coding practices: Tests whether the model generates code with security vulnerabilities, based on the CWE standard.
• Prompt injection: Evaluates the model's ability to identify untrusted input and resist attacks.
• Compliance with cyberattack requests: Measures the model's response rate to requests with malicious intent and its False Refusal rate.
• Code interpreter abuse: Tests resistance to attacks that attempt to induce the execution of malicious code within a sandboxed environment.
• Automated attack capability: Evaluates the ability to perform real-world exploits such as SQL injection through CTF-style tests.

Key Findings:

• Industry-wide security improvement: Since the first version was released in December 2023, the average model compliance rate with cyberattack requests has decreased from 52% to 28%, indicating improved security.
• Narrowing gap between models: Code-specialized models had tended to be less secure than general-purpose models, but this gap has recently narrowed as security capabilities rapidly catch up.