Atlanta Cybersecurity: Team Atlanta Takes AI Open Source

The Atlanta cybersecurity project developed through the Defense Advanced Research Projects Agency’s Artificial Intelligence Cyber Challenge, commonly called AIxCC. DARPA launched the two-year competition in 2023 to encourage the development of AI systems that could find and repair vulnerabilities in open-source software.

Team Atlanta created a cyber reasoning system called Atlantis for the competition. A cyber reasoning system analyzes software, tests possible weaknesses, and develops patches with limited human direction.

The team brought together researchers and engineers from Georgia Tech, the Georgia Tech Research Institute, Samsung Research, Samsung Research America, the Korea Advanced Institute of Science and Technology, and Pohang University of Science and Technology. The participating groups combined experience in information security, computer engineering, artificial intelligence, fuzzing, and vulnerability research.

Team Atlanta won the final AIxCC competition at DEF CON in August 2025 and received the $4 million first-place prize. The final round tested seven systems on large open-source projects and required them to identify vulnerabilities, prove that the flaws could affect software, and produce possible repairs. Georgia Tech reported that the victory followed two years of research and engineering work.

Atlantis combines large language models with established software analysis methods to increase automation during vulnerability testing. The system uses AI to understand large codebases while security tools test whether a suspected flaw can cause an actual failure.

One of those methods is fuzzing. A fuzzer repeatedly sends unusual or unexpected inputs to a program. Security researchers then study any crash or abnormal response that the test produces. Team Atlanta also used symbolic and concolic execution, which examine different paths that a program may follow when it processes information.

Large language models helped the system apply these methods to large software projects. In an OpenSSF interview, Georgia Tech professor Taesoo Kim explained that Team Atlanta used language models to support traditional analysis tools. The models helped the system understand input formats, generate testing vocabulary, and focus its analysis on relevant parts of a codebase.

A suspected flaw does not automatically qualify as a proven vulnerability. The system must create an input that reaches the affected code and produces the harmful result. This proof helps researchers separate valid findings from false alarms.

Software maintainers often receive more security reports than they can investigate. A system that provides clear evidence and a possible repair can help them decide which findings require immediate attention.

AIxCC required the finalists to release their cyber reasoning systems as open-source software. However, publishing the code did not make each system easy to operate outside the competition.

Many systems depended on the specialized infrastructure that DARPA used during AIxCC. Developers could review the code after the competition, but they faced technical barriers when they tried to run it on other projects or computing environments.

Georgia Tech’s Systems Software Lab began developing the Open Source Cyber Reasoning System framework, or OSS-CRS, to address these barriers. Professor Kim directs the lab, while Georgia Tech doctoral student Andrew Chin leads the OSS-CRS program.

OSS-CRS provides a common structure for developing and operating AI-based security systems. Team Atlanta has moved Atlantis into the framework, which allows developers to run it with different resource and budget settings. According to the Georgia Tech announcement, the team designed the new version to work on laptops and other commonly available machines.

The Open Source Security Foundation accepted OSS-CRS as a sandbox project within its AI/ML Security Working Group. OpenSSF operates under the Linux Foundation and supports projects that improve the security of open-source software.

The sandbox status gives OSS-CRS access to technical guidance, working group meetings, industry outreach, and connections with software maintainers. These resources can help the project develop through open collaboration instead of depending on one research team.

OSS-CRS serves as an orchestration framework. It creates a shared operating structure for different cyber reasoning systems instead of limiting developers to one security model.

The framework provides several main functions:

• A standard system interface. Developers can build a cyber reasoning system once and operate it across supported local and cloud environments.
• Compatibility with OSS-Fuzz projects. OSS-CRS can test projects that already use the OSS-Fuzz format, which supports automated vulnerability testing.
• Support for multiple security systems. Developers can combine several cyber reasoning systems during one analysis campaign.
• Control over computing resources. Teams can set processor limits and language-model budgets for each system.

These controls can help researchers compare different approaches and manage the cost of large software analyses. A smaller project may run one system with limited resources, while a larger campaign may combine several systems for wider coverage.

OSS-CRS also allows Atlantis to work with systems created by other AIxCC finalists and newer command-line security agents. Each system may use different methods, so a combined analysis can find weaknesses that one tool might miss.

The open framework gives universities, software companies, and independent security researchers across the United States and other countries a shared starting point. Contributors can test the software, report problems, propose improvements, and study how different AI security methods perform. This open process can improve transparency by allowing researchers to examine the framework’s methods and limitations.

AI can help researchers analyze more code, but automated testing does not guarantee that every proposed patch is correct. A repair may pass a program’s tests while changing its intended behavior or creating another problem.

The OSS-CRS researchers manually reviewed 630 AI-generated patches during one study. They found that 20% to 40% of the patches were semantically incorrect, even though the patches passed the automated validation process. A semantically incorrect patch may compile and pass tests while failing to preserve what the software should do.

These findings support the need for human oversight. AI governance requires accuracy, accountability, and clear review procedures, especially when automated systems influence security decisions.

A responsible review process should include several actions:

• Confirm the vulnerability. Reviewers should reproduce the issue and determine whether an attacker could use it.
• Check the project’s threat model. A reported behavior may fall outside the security conditions that a project accepts.
• Examine the proposed patch. Maintainers should confirm that the repair preserves the software’s intended behavior.
• Coordinate the disclosure. Security teams should give maintainers time to review and repair a flaw before publishing sensitive details.

Human review also protects software maintainers from low-quality reports. Automated systems can create large numbers of findings, and each report takes time to assess. Verification helps ensure that maintainers receive evidence they can use.

OpenSSF plans to connect cyber reasoning projects with open-source communities. These connections can support vulnerability triage, patch review, and responsible disclosure before a report reaches the public.

OSS-CRS has already tested software outside the AIxCC competition. OpenSSF reported in May 2026 that Team Atlanta had used the framework to find 25 vulnerabilities across 16 open-source projects.

The affected software included PHP, U-Boot, memcached, and Apache Ignite 3. At the time of the report, maintainers had fixed nine vulnerabilities and confirmed eight more with repairs in progress. These figures came from the OpenSSF review of AIxCC’s results.

The results show that cyber reasoning systems can contribute to active software projects. They also show why researchers must work with maintainers. A tool can identify suspicious code and propose a patch, but project experts still need to assess the finding within the software’s actual purpose and security requirements.

AIxCC systems found real weaknesses during the competition as well. Independent engineers reproduced 27 candidate issues after several rounds of testing. Some findings required repairs, while others fell outside a project’s threat model or had already been addressed through separate code changes.

The review process helped prevent inaccurate reports from reaching maintainers. It also gave researchers more information about how cyber reasoning systems should assess their own findings before submitting them.

The Georgia Tech cybersecurity project connects research at the Georgia Institute of Technology with an international open-source community. Researchers in Atlanta can continue developing the system, while software maintainers and security specialists in other locations can test and improve it.

Team Atlanta’s structure also shows how academic and industry researchers can share technical knowledge. Georgia Tech contributed students and security research, while Samsung and the participating Korean universities added engineering and vulnerability expertise.

Atlanta already supports partnerships between academic laboratories and corporate research teams. Peach State Tech has reported on how companies work with researchers through Atlanta’s Tech Square. OSS-CRS adds an open-source model to that local research activity because contributors outside the original partnership can participate.

Georgia Tech also helps convene wider discussions about the responsible use of artificial intelligence. Its involvement in the ACM AI Leadership Summit Atlanta connects researchers with business leaders, government officials, and other technology professionals.

OSS-CRS gives these discussions a practical example. The project demonstrates how researchers can combine AI with established testing methods while retaining human responsibility for final security decisions.

Cyber reasoning systems still face technical and operational limits. Large software projects may contain millions of lines of code, several programming languages, complex dependencies, and incomplete documentation. Each condition can make automated analysis harder.

Computing costs also affect access. Language models and large testing campaigns can require significant processing capacity. OSS-CRS includes budget and resource controls, but project teams still need to decide how much computing time each analysis deserves.

False positives create another concern. A system that reports too many weak findings can increase the workload of open-source maintainers. Security projects need strong validation procedures so they can submit fewer reports with clearer evidence.

AI-generated patches create a separate risk. A patch may stop one test from failing while damaging another function. Human reviewers must examine the code, run additional tests, and confirm that the repair matches the project’s intended behavior.

Open-source development gives the community a way to study these limits. Researchers can compare systems, publish findings, and improve validation methods as more developers use OSS-CRS. Organizations of all sizes depend on open-source components, including many that operate inside commercial products and public services. A vulnerability in one widely used component can therefore affect many separate systems.

OpenSSF now provides a neutral organization for continued OSS-CRS development. Contributors can work through its AI/ML Security Working Group and Cyber Reasoning Systems Special Interest Group.

The project’s next phase will depend on testing, documentation, maintenance, and cooperation with open-source projects. Developers need clear instructions for installing the framework, configuring its resources, evaluating findings, and reporting vulnerabilities responsibly.

Researchers must also improve the quality of AI-generated patches. The ensemble feature offers one possible method. OSS-CRS can compare patches from multiple systems and select the strongest candidate. Early research found that this combined process could equal or exceed the performance of individual systems.

The project may eventually help maintainers find software flaws earlier, reduce the time required to prepare repairs, and improve the resilience of widely used open-source software. Its progress will depend on careful testing and continued human review.

Peach State Tech reports on the founders, researchers, companies, and institutions shaping Georgia’s technology economy. Follow Peach State Tech for clear coverage of Georgia Tech cybersecurity research and other developments influencing the future of Atlanta cybersecurity.