Cisco, former Google, Meta experts train cybersecurity LLM

Agentic AI for cybersecurity: Panacea or Pandora's box?

In March, Trend Micro released an AI reasoning model based on Meta's Llama 3, which was trained on cybersecurity data and has open weights. Increasing specialization of LLMs is an expected evolution now that initial foundation models have been established and reached limited returns for specialized tasks, said Adrian Sanabria, an independent security consultant.

"There is a lot of room for innovation in creating specialized models, without necessarily making any of the models more powerful," Sanabria said. "Reasoning models in an agentic architecture will route tasks to the most appropriate model, API, or service. We're seeing new standards, protocols and architecture emerge to handle all this, [such as] Google's Agent2Agent protocol and [Model Context Protocol] MCP servers."

However, he said, even specialized AI agents that use cybersecurity LLMs are likely to encounter scalability challenges as their use grows.

First, there's the cost -- even a single A100 costs about $8,000, and agentic AI consumes more energy than traditional IT automation workloads, Sanabria said, citing an April 21 analysis by Tim MalcomVetter, CEO and co-founder of Wirespeed, a managed detection and response vendor.

Meanwhile, another security operations startup, Panther, estimates the average SOC receives more than 4,000 alerts per day.

"That's 167 alerts to process per hour. Dropzone AI states it takes AI SOC agents 3 to 11 minutes per alert," Sanabria said. "Cisco would have to be processing 2.7 alerts per minute to keep up with the average."

That's still much faster than the average human SOC analyst, at 20 to 40 minutes, according to Dropzone, but "if each alert is taking Dropzone AI 3 minutes in the best-case scenario, that's a limit of 480 alerts per day, assuming their estimate is for a single GPU," Sanabria said.

Ultimately, agentic AI alone won't alleviate SOC alert fatigue, Sanabria said.

"The answer is to be more selective about the alerts you use AI on, because it's expensive and limited," he said. Existing AI-based tools can also perform event correlation and alert reduction -- but don't need LLMs.

The answer is to be more selective about the alerts you use AI on, because it's expensive and limited.

"There are [managed security service providers] automating this at scale without LLMs," he said. "The LLM use seems to be largely driven by hype. My prediction is that if using GenAI to triage alerts and recommend actions works, we'll start seeing acquisitions later this year."

Releasing a cybersecurity LLM to open source also carries potential risks, according to Thurai.

"There are some concerns that bad players can use LLMs like this to scan for vulnerabilities in the enterprise and use it for attacks," he said. "There's always the possibility that if they open source something like this, adversaries can use it to attack things that much faster."

Still, if it performs as advertised, the Cisco Foundation AI model will mostly help enterprises identify vulnerabilities, Thurai said, including during red-teaming exercises.

"There are also already open source tools available that are used by white hats within enterprises," he said.

* Editor's Note: Cisco clarified after publication that it recruited engineers from Meta and Google, rather than leading a collaboration between companies, as was originally reported.

Beth Pariseau, a senior news writer for Informa TechTarget, is an award-winning veteran of IT journalism covering DevOps. Have a tip? Email her or reach out @PariseauTT.