Cisco Updates LLM Security Leaderboard with Multi-Modality Evaluations

The Cisco LLM Security Leaderboard has been updated to include evaluations across multiple modalities, including text, image, and audio. This comprehensive leaderboard aims to provide organizations with clear data on how models hold up against attacks before deployment.

With the rise of AI-powered agents that can read email, browse the web, and take actions on behalf of users, there is a growing risk of manipulation by malicious instructions or prompts. The leaderboard tests for this risk using various techniques, including prompt injection, jailbreaks, and other methods that push models toward harmful output.

Since June 2026, 102 new evaluations have been added across the three modalities, bringing the total number of models on the leaderboard to 136. These models come from various labs, including Anthropic, OpenAI, Google, Meta, Mistral, and xAI.