Achieved AA Rating in ATLA's 5-Year Research (Type S) - Results Exceeding Expectations with Reinforcement Learning Fuzzing Technology
Ricerca Security obtained an 'AA' rating in an ATLA research project for developing a fuzzing technology using reinforcement learning that discovered 26 zero-day vulnerabilities.
📋 Article Processing Timeline
- 📰 Published: April 14, 2026 at 22:19
- 🔍 Collected: April 14, 2026 at 14:01
- 🤖 AI Analyzed: April 19, 2026 at 16:25 (122h 23m after Collected)
Ricerca Security, Inc. announces that its 5-year research project conducted under the "National Security Technology Research Promotion System" of the Acquisition, Technology & Logistics Agency (ATLA) has concluded, receiving a final evaluation of "AA (results exceeding expectations)."
In this research, we developed vulnerability detection technology utilizing reinforcement learning and achieved results in both academic and practical aspects, such as discovering numerous zero-day vulnerabilities from actual software. This achievement is expected to contribute to the advancement of vulnerability discovery in the cybersecurity field and the improvement of critical infrastructure safety.
Received AA evaluation in ATLA's large-scale research system (Type S)
Ricerca Security, Inc.'s research project "Proposal of an Environment-Adaptive Fuzzing System using Reinforcement Learning" (FY2020-FY2024), adopted as a Type S (large-scale, long-term research framework) under the "National Security Technology Research Promotion System" implemented by ATLA, has concluded and received a final evaluation of "AA (results exceeding expectations)."
This system is a competitive research framework aimed at creating advanced technologies that contribute to national security, and Type S, in particular, targets long-term and challenging research and development.
Background: Increasingly sophisticated cyber attacks and the limits of vulnerability discovery
In recent years, software vulnerabilities have become a serious issue affecting national security and social infrastructure. In particular, "zero-day vulnerabilities," for which no patches exist, tend to cause widespread damage and require rapid discovery.
On the other hand, traditional vulnerability discovery methods had the following challenges:
- Difficulty in selecting the optimal search method
- Difficulty in determining the severity of a large number of discovered defects
- Difficulty in application to real-world environments such as IoT devices
Research Results: Establishment of next-generation fuzzing technology using reinforcement learning
In this research, we developed fuzzing technology and related infrastructure utilizing reinforcement learning with the aim of improving search efficiency and expanding the scope of application in fuzzing.
The main results are as follows:
Development of the integrated fuzzing framework "fuzzuf"
We developed "fuzzuf," a framework capable of handling multiple fuzzing algorithms in an integrated manner.
In this framework, it is possible to execute, compare, and combine 13 types of existing fuzzing algorithms (AFL, libFuzzer, VUzzer, etc.) on the same platform.
Related: Cybersecurity startup Ricerca Security open-sources the domestic fuzzing framework "fuzzuf," contributing to streamlined vulnerability detection.
Investigation of fuzzing optimization methods using reinforcement learning
To improve search efficiency in fuzzing, we investigated optimization methods using reinforcement learning.
Specifically, we proposed and implemented:
- A method to optimize the application order and frequency of mutation operations (SLOPT)
- A method to dynamically switch among multiple fuzzing algorithms
And evaluated the performance of each method.
Vulnerability discovery in actual software
As a result of applying the proposed methods to actual software, we discovered:
- A total of 26 zero-day vulnerabilities
- 17 of them registered as CVEs
- 5 evaluated as critical vulnerabilities
Thereby confirming its effectiveness in real-world environments.
Development of crash cause analysis and threat evaluation methods
For the large number of crashes obtained through fuzzing, we performed:
- Development of a comparison and verification platform (RCABench) for methods identifying the root cause location
- Design of metrics to evaluate the threat level of vulnerabilities
As a result of the evaluation, we confirmed that threat evaluation of known vulnerabilities is possible with an accuracy equal to or higher than manual analysis.
Extension of fuzzing methods for IoT devices
For environments where internal structures are difficult to obtain, such as IoT devices, we proposed a "coverage estimation method combining static analysis results and communication responses (Shepherd)" and confirmed an improvement in accuracy compared to existing methods.
Evaluation Results: Evaluated as "results exceeding expectations"
In the final evaluation by ATLA, this research was rated "AA (results exceeding expectations)" as a comprehensive evaluation.
In the evaluation, the following achievements were confirmed:
- Construction of a framework capable of integrating multiple fuzzing algorithms
- Reinforcement learning...
In this research, we developed vulnerability detection technology utilizing reinforcement learning and achieved results in both academic and practical aspects, such as discovering numerous zero-day vulnerabilities from actual software. This achievement is expected to contribute to the advancement of vulnerability discovery in the cybersecurity field and the improvement of critical infrastructure safety.
Received AA evaluation in ATLA's large-scale research system (Type S)
Ricerca Security, Inc.'s research project "Proposal of an Environment-Adaptive Fuzzing System using Reinforcement Learning" (FY2020-FY2024), adopted as a Type S (large-scale, long-term research framework) under the "National Security Technology Research Promotion System" implemented by ATLA, has concluded and received a final evaluation of "AA (results exceeding expectations)."
This system is a competitive research framework aimed at creating advanced technologies that contribute to national security, and Type S, in particular, targets long-term and challenging research and development.
Background: Increasingly sophisticated cyber attacks and the limits of vulnerability discovery
In recent years, software vulnerabilities have become a serious issue affecting national security and social infrastructure. In particular, "zero-day vulnerabilities," for which no patches exist, tend to cause widespread damage and require rapid discovery.
On the other hand, traditional vulnerability discovery methods had the following challenges:
- Difficulty in selecting the optimal search method
- Difficulty in determining the severity of a large number of discovered defects
- Difficulty in application to real-world environments such as IoT devices
Research Results: Establishment of next-generation fuzzing technology using reinforcement learning
In this research, we developed fuzzing technology and related infrastructure utilizing reinforcement learning with the aim of improving search efficiency and expanding the scope of application in fuzzing.
The main results are as follows:
Development of the integrated fuzzing framework "fuzzuf"
We developed "fuzzuf," a framework capable of handling multiple fuzzing algorithms in an integrated manner.
In this framework, it is possible to execute, compare, and combine 13 types of existing fuzzing algorithms (AFL, libFuzzer, VUzzer, etc.) on the same platform.
Related: Cybersecurity startup Ricerca Security open-sources the domestic fuzzing framework "fuzzuf," contributing to streamlined vulnerability detection.
Investigation of fuzzing optimization methods using reinforcement learning
To improve search efficiency in fuzzing, we investigated optimization methods using reinforcement learning.
Specifically, we proposed and implemented:
- A method to optimize the application order and frequency of mutation operations (SLOPT)
- A method to dynamically switch among multiple fuzzing algorithms
And evaluated the performance of each method.
Vulnerability discovery in actual software
As a result of applying the proposed methods to actual software, we discovered:
- A total of 26 zero-day vulnerabilities
- 17 of them registered as CVEs
- 5 evaluated as critical vulnerabilities
Thereby confirming its effectiveness in real-world environments.
Development of crash cause analysis and threat evaluation methods
For the large number of crashes obtained through fuzzing, we performed:
- Development of a comparison and verification platform (RCABench) for methods identifying the root cause location
- Design of metrics to evaluate the threat level of vulnerabilities
As a result of the evaluation, we confirmed that threat evaluation of known vulnerabilities is possible with an accuracy equal to or higher than manual analysis.
Extension of fuzzing methods for IoT devices
For environments where internal structures are difficult to obtain, such as IoT devices, we proposed a "coverage estimation method combining static analysis results and communication responses (Shepherd)" and confirmed an improvement in accuracy compared to existing methods.
Evaluation Results: Evaluated as "results exceeding expectations"
In the final evaluation by ATLA, this research was rated "AA (results exceeding expectations)" as a comprehensive evaluation.
In the evaluation, the following achievements were confirmed:
- Construction of a framework capable of integrating multiple fuzzing algorithms
- Reinforcement learning...