Skip to content

Response, New York. Noah Park: Check Point Patches Actively Exploited Management Server Path Traversal. Detection, Dallas. Ava Okello: Report Finds Nearly Half of Deployed Detections Need Attention Before SOCs Can Trust Them.

SOCtember

Always First. Fast SOC News.

Kaspersky chart comparing three filter generation approaches, Algorithm, LLM and Agents, with approval rates of 0.61, 0.94 and 0.96 and false positive drop of about 0.85, about 0.99 and about 0.99.
Image: Kaspersky (fair use). Comparison of the three approaches from Kaspersky's Oct. 30, 2026, Securelist post.

Tools

Kaspersky Tested Three Ways to Write False Positive Filters for SOC Alerts and Picked a Single LLM Prompt Over a Four-Agent Pipeline, Saying It Scored Almost the Same and Runs Faster

In Kaspersky's comparison, analysts approved an algorithmic generator's filters at a rate of 0.61, a single-prompt LLM's at 0.94 and a chain of four agents' at 0.96. The company says the LLM approach it chose cut false positives by 99%, though it did not name the model or say how large the evaluation was.

James Whitford, ToolsAustin5 min read

AUSTIN - Kaspersky has tested three ways to automate one of the most tedious jobs in a security operations center, writing the filters that suppress false positive alerts, and settled on a single large language model prompt over a four-agent pipeline because the two scored almost the same and the simpler one is faster, the company wrote on its Securelist blog on Friday.

The post compares a purely algorithmic generator, an LLM working from one system prompt and a chain of LLM agents. Kaspersky said the LLM approach it chose cut false positives by 99%, increased the number of filters generated per day by 28% and made analysts' alert handling 17% more efficient. It did not say over what period or across how many alerts those figures were measured.

The problem it is trying to solve is familiar to anyone who has tuned a SIEM. A filter, in Kaspersky's definition, is a combination of event field values that marks a detection as a false positive. Get it too broad, such as a filter on PowerShell alone, and it hides alerts that deserve attention. Get it too narrow, such as a filter on a single event ID, and it covers one event that may never recur. Events in its pipeline carry around 400 fields on average, and four may be enough for a filter.

Three approaches, one workflow

Kaspersky diagram of the filter generation workflow, showing an alert passing through candidate selection to generation, which outputs filters and a description.
Image: Kaspersky (fair use).The filter generation workflow from the Securelist post.

In every version, the trigger is the same: an analyst decides an alert is a false positive and asks the generator for help. The system takes the 400-field event and the rule that fired, and offers three candidate filters.

The algorithmic version first determines the rule fragment's context, the set of events that fragment responds to, then selects relevant fields, builds combinations with the FP Growth algorithm and ranks them. Kaspersky said its strengths are that it is deterministic, interpretable and has no external model dependencies, so it can run in any customer environment. Its weaknesses are that it cannot create a filter when there is no context, such as a rule's first-ever trigger, and that normalizing strings like file paths across systems runs into atypical cases the rules do not cover.

The single-prompt LLM version keeps the algorithm for field selection, which Kaspersky said a simple algorithm handles without errors, then hands the event, the selected fields and the rule to the model, which returns three filters and an explanation. That removed the dependence on context and handled normalization more flexibly, the company said, but it had no output validation, and because one system prompt handled every task, "a minor fix in one place could break the logic in another."

Kaspersky illustration of three alerts with candidate filter fields such as file_path, process_name, netremoteipv4 and parentprocessfilepath, the first two alerts marked filter chosen and the third marked nothing relevant.
Image: Kaspersky (fair use).Example candidate filters and analyst choices from the Securelist post.

The agent version splits the job across four agents: a Selector that chooses relevant fields, a Generator that writes filters, a Validator that sends invalid filters back for regeneration or restarts field selection, and a Commentator that explains which field values were used and why. It can validate and correct its own output, but it makes more model calls, so it is slower, and it depends on access to a model that may not be available in an isolated environment.

How Kaspersky measured it

Kaspersky used two metrics. The approval rate counts how often the analyst accepts the suggestions for an alert. The false positive drop compares false positive alerts in a set period before and after a filter is deployed, so it can only be measured over time.

In Kaspersky's comparison, the algorithm scored an approval rate of 0.61 with a false positive drop of about 0.85. The single-prompt LLM scored 0.94 and about 0.99. The agents scored 0.96 and about 0.99. "So we chose the simpler and faster option: an LLM with a single system prompt," the post said.

Kaspersky example JSON event used for filter generation, with the file_path of powershell.exe and a filecmdline running Get-VM -Name kafka underlined.
Image: Kaspersky (fair use).An example event from the Securelist post.

What SOC teams can take from it

The post is a vendor describing its own system, and it does not name the model, publish its prompts or say how large the evaluation was. But the design choices map onto problems any detection team will recognize. Kaspersky removes the fields a detection rule already matches on, because a filter that mirrors the rule's own query "effectively disables that rule." It normalizes paths by replacing drive letters and user folders with standard environment variables and converting to one case. And it keeps a human in the loop. In the algorithm and single-prompt examples, two of the three suggested filters were valid and one was too broad; in the agent example, all three were valid. The analyst still picks the filter, and in one example edited it before use.

For teams without access to a model in their environment, Kaspersky's own conclusion is that the algorithm is still useful. "In two-thirds of cases, it still produces filters that satisfy the analyst," the company wrote, and it works best when enough similar events have accumulated.

About the Author

James Whitford

James Whitford covers SOC tooling, SIEM, SOAR, and detection platforms for SOCtember from Austin.

Related stories

Tools desk