You probably remember that this summer our red team published a research on using LLMs to generate offensive tooling. The experiment began with a failed attempt to solve the task with a one-shot prompt; however, a successfully working pipeline of three AI agents (Orchestrator, Tester, Developer) was then built, which quite rapidly generated functional Mythic agents.
At the same time, our blue team experts conducted experiments on using LLMs for defense: to generate filters that should filter out false positives from the stream of security systems’ alerts. This study also compared various approaches: a purely algorithmic pipeline (FP-Growth), a single LLM-based generator, and a generator built on a group of LLM agents (Selector, Generator, Validator, Commentator).
The results showed that the simple algorithmic approach is not so bad: although it scored lower on the evaluation metrics, but in two-thirds of cases, it still produces filters that satisfy the analyst. The advantages of this method are determinism and interpretability, as well as independence from external models; it can be deployed in any environment.
Both standalone LLM and multi-agent LLM approaches show high accuracy, and their results differ only slightly, while the group approach is slower. If an LLM is available, the best option appears to be a single model that handles several stages of the generation cycle.
Comparing these results with the aforementioned “red team” study, we assume that LLM agents will be used more often in attacks (and in offensive security) than in defense. The main reason for that: defenders cannot afford to make mistakes and prefer more deterministic technologies.
However, this is merely our assumption. For the details of the alert filtering research, see the Artemy Obukhov’s article “Filter generation in practice: from algorithms to LLMs”.
