INTELLIGENT BRANDS // Enterprise Security

New AI SOC LLM Leaderboard measures LLMs in realistic IT environment.

The new benchmark compares LLMs across a diverse range of attacks and SOC tools in a realistic IT environment over all phases of alert investigation, from alert ingestion to disposition and reporting. It includes a public leaderboard to help professionals decide the best LLM for their SOC needs.

“ Our industry-first benchmark enables SOC teams and vendors to pick the best LLM for this purpose. This benchmark is made possible by Simbian’ s AI SOC Agent – a proven solution leading the industry in endto-end alert investigation leveraging LLMs.”

Existing benchmarks compare LLMs over broad criteria such as language understanding, math, and reasoning. Some benchmarks exist for broad security tasks or very basic SOC tasks like alert summarization.

But prior to Simbian’ s announcement, no benchmark existed to comprehensively measure LLMs on the primary role of SOCs, which is to investigate alerts end-to-end. This task involves diverse skills, including the ability to:

• Understand data, extract evidence, and map it to attack stages

• Reason over evidence to arrive at a clear disposition and severity

• Produce clear reports and response actions

• Customize investigations for each organization’ s context.

To make the benchmark applicable across a range of SOC environments, AI SOC LLM Leaderboard leverages 100 diverse full-kill chain scenarios that test all layers of defense. It is also the industry’ s first benchmark to measure investigation performance in a lab environment mimicking an enterprise, with investigations autonomously retrieving data from live tools across the environment.

“ SOC analysts and vendors building tools for the SOC are rapidly embracing LLMs to scale their operations, increase accuracy and reduce costs,” said Ambuj Kumar, Simbian CEO and Co-Founder.

• Understand alerts from a broad range of detection sources

• Determine how to investigate any given alert

• Generate code to support that investigation

This first LLM benchmark tested today’ s top-tier LLM models from Anthropic, OpenAI, Google, and DeepSeek. All tested models were able to complete over half( 61 %-67%) of the tasks involved in alert investigation, as long as there was a solid framework to break down an investigation into clearly defined tasks for LLMs.

The AI SOC LLM Leaderboard reveals that LLMs are more capable than commonly believed for autonomous alert investigation. Marginal difference was observed between standard LLMs and thinking LLMs for alert investigation. The results showed that the best LLM for cybersecurity is a generalist( like Sonnet 3.5) that knows how to code as well as how to perform logical reasoning, rather than a specialist that excels at code( Sonnet 4.0) or at logical reasoning( Opus 4). Finally, the benchmark highlighted that specialization such as SOC-specific training or a mix of LLMs yields higher performance than any single LLM. p

68 INTELLIGENTCIO NORTH AMERICA www. intelligentcio. com