LLM × Agent × VM ─ Malware Analysis Automation

Complex reports, fully automated.

r_AKanaly is a malware analysis automation tool. An LLM and custom agents operate a VM, cross-check dynamic analysis logs with Ghidra static analysis data, and produce a grounded report.

Mission

What r_AKanaly aims for

Malware analysis reports are essential for recovering from damage and stopping the same attack again, yet writing one takes analysis knowledge and skill with specialized tools. For beginners that is a high barrier, and even experts need time for every single report, so they cannot produce them in large numbers, and the report you are looking for often simply does not exist. r_AKanaly aims to increase the number of malware analysis reports available to everyone, making "create a report" as easy an option as uploading a sample or looking up its hash. To get there, the goal is a system where, once the initial environment is set up, you only specify the sample and the required information, and everything from analysis to saving the report and artifacts and shutting down the VM completes fully automatically.

Sample r_AKanaly Reports worldwide

01 ─ Overview

What is r_AKanaly?

Hand it a sample, and the LLM decides how to investigate while custom agents operate a VM to collect data. It then writes up findings that cross-check what was observed against the code.
It removes the burden of operating the VM by hand at every step, connecting analysis all the way to the report.

Sample

The file to analyze and
the analysis conditions

Automated analysis

An LLM and custom agents
operate the VM

Grounded report

Facts, inferences, evidence,
and open questions

Decide

The LLM decides what to investigate next, reads the collected data, and cross-checks and reasons over it.

Operate

Custom agents turn the LLM's instructions into concrete operations and data retrieval inside the VM, then return the results.

Write

Findings in the report are tied to the observed records and code data that were actually collected.

Strength 01 Autonomous VM Operation

Fully automated VM operation
by an LLM and custom agents

The LLM is used not just to write text, but as the decision-maker that drives the analysis. Custom agents turn its instructions into operations and data retrieval inside the VM, and return the results to the LLM. Scroll to follow one full loop.

Host Guest (VM) LLM MCP Host agent OS agent Deploy / prep / OS ops Ghidra agent Functions & analysis Sample Results Next decision
  1. 1

    The LLM decides the next investigation

    Based on the sample, the analysis conditions, and everything learned so far, the LLM decides "what should be checked next."

  2. 2

    A tool is called through MCP

    MCP is the gateway the LLM uses to call tools. Its decision is sent as a tool call.

  3. 3

    The Host agent receives it and sends commands to each agent

    Outside the VM (on the Host), the Host agent receives the instruction and, depending on what it asks for, sends commands to the OS agent and Ghidra agent inside the VM.

  4. 4

    Agents inside the Guest do the work

    Inside the VM (the Guest), the OS agent handles deployment, preparation, and OS operations, while the Ghidra agent retrieves functions and other analysis data.

  5. 5

    Results return to the LLM

    The collected records and analysis data go back to the LLM through the Host. Nobody has to operate the VM by hand and copy things over.

  6. 6

    Read the results, decide again

    The LLM reads the results and decides whether more investigation is needed. Repeating this loop carries the analysis all the way to the report.

Manual analysis

At every step, a person operates the VM, pulls out the results, reads them, and thinks about the next move.

r_AKanaly

The LLM decides, custom agents operate. Decide → operate → results → decide runs on its own.

Strength 02 Log Compression for Reasoning

Compressing dynamic logs
for reasoning

Running a sample produces a huge amount of dynamic analysis logs. r_AKanaly organizes them into behavior data that keeps what the analysis needs, and uses it for the LLM's reasoning.

Raw dynamic analysis log 0 lines

Too much to reason over as-is

Organized behavior data example
  • Process

    Spawned a child process and ran a command

  • File

    Wrote an executable to a temp folder

  • Registry

    Set a value on an autorun-related key

  • Network

    Tried to connect to an external host

Grouped into units the LLM can easily understand

"Compression" here does not mean ZIP-style compression

File compression (ZIP, etc.)

  • Makes the file smaller
  • The contents are the same records
  • Must be extracted to read

r_AKanaly log compression

  • Organizes while keeping what the analysis needs
  • Groups records into units of behavior
  • Feeds straight into the LLM's reasoning

The compressed behavior data is used in two ways

Compressed behavior data
A

Understand what the sample did

What it ran, what it wrote, and where it tried to connect.

B

Decide where to look in static analysis

Focus Ghidra on the code that is likely related to the observed behavior.

Strength 03 Grounded Report

Grounded reports

Grounding means tying explanations and inferences to data that was actually collected. You can trace each statement in the report back to the observed records and analysis data it came from.
Hover over (or tap) a statement in the report below to see its evidence.

Analysis report Example (fictional data)

Findings

  • Fact After execution, the sample spawned a child process and wrote a file to a temp folder.
  • Fact A write to an autorun-related registry key was recorded.
  • Inference The registry operation in function FUN_00401a30 likely corresponds to the observed write.
  • Inference Since it is called from the initialization routine, it is likely meant to keep running after a reboot.
  • Open question The routine that decides the connection target was not identified in this analysis. Flagged for further investigation.

Dynamic analysis records

  • L1Process Create parent=sample.exe child=cmd.exe
  • L2File Write %TEMP%\a3f1.tmp
  • L3Reg SetValue HKCU\…\CurrentVersion\Run\updater
  • L4TCP Connect → 203.0.113.10:443 (failed)

Ghidra analysis data

  • G1FUN_00401a30 : calls RegSetValueExW
  • G2Caller of FUN_00401a30 : FUN_00401200 (init)
  • G3FUN_00402c10 : builds a string for the connect routine (arg unresolved)
Fact

What was actually confirmed by dynamic analysis or Ghidra.

Inference

Conclusions drawn by cross-checking facts against code data.

Evidence

Which records and analysis data each statement is based on.

Open question

What could not be confirmed and needs more investigation.

Separating facts from inferences and keeping the evidence traceable makes it possible to verify the results and investigate further.

03 ─ Architecture

How it works

The Host side controls the workflow and passes instructions and results, while dynamic analysis and Ghidra static analysis run inside the Guest VM. Click any element to see its role.

Decision
Gateway
Host
Guest (VM)

Division of roles

LLM

Decides what to investigate, reads analysis data, cross-checks, reasons, and writes the report content

Custom agents

Turn instructions into concrete operations and data retrieval, and return the results

Scripts

Run fixed routines, such as standard processing and connecting steps

VM and analysis tools

The environment and means for running and observing the sample and retrieving code and function data

04 ─ Workflow

Analysis workflow

The analysis moves through six stages. Because investigation includes fetching and checking more data as needed, it is not a fixed, one-pass pipeline.

  1. 1

    Specify the sample and conditions

    Provide the target file and the premises of the investigation.

  2. 2

    Observe behavior with dynamic analysis

    Run it inside the VM and collect records such as process and file activity.

  3. 3

    Compress and organize the logs

    Turn the huge volume of records into behavior data the LLM can reason over.

  4. 4

    Examine the internals with static analysis

    Use Ghidra functions and analysis data to find the code related to the observed behavior.

  5. 5

    Cross-check and reason

    Connect the dynamic observations with the code to work out what happened and which routines were involved.

  6. 6

    Write a grounded report

    Organize facts, inferences, evidence, and open questions.

05 ─ Scope

Scope of automation

Once the user has prepared the analysis environment, the sample, and the conditions, r_AKanaly connects everything that follows automatically.

Prepared by the user

  • Analysis environment
  • Sample
  • Analysis conditions

Connected automatically by r_AKanaly

  1. VM operation
  2. Data retrieval
  3. Log organizing & compression
  4. Cross-checking & reasoning
  5. Report generation

* Initial environment setup and specifying conditions are treated separately from the automated analysis that follows.

06 ─ Input & Output

Input and output

What you hand over, what is produced along the way, and what you end up with.

Input

  • The sample to analyze
  • Analysis conditions
  • Analysis environment settings

Produced along the way

  • Dynamic analysis logs
  • Compressed behavior data
  • Functions and static analysis data from Ghidra
  • Mapping between observations and code

Final output

Analysis report
  1. Target and conditions
  2. Observed behavior
  3. Static analysis findings
  4. Evidence-based findings
  5. Open questions
See real reports (Japanese) →

GET r_AKanaly

Start analyzing.

Choose your OS and PC type, then download.

Read the README first

Requirements, setup steps, and cautions are in the README of the GitHub repository.

OS
PC type
Download