What happened
According to ChosunBiz, Kim Tae-su said specialized AI agents working together can improve vulnerability discovery, verification, exploit testing, and patch generation. The report cites Team Atlanta’s performance in DARPA’s AI Cyber Challenge and describes Microsoft’s MDASH system, which uses more than 100 sub-agents to analyze software.
ChosunBiz reported that Kim Tae-su, Microsoft’s corporate vice president for security and a Georgia Institute of Technology computer-engineering professor, made the comments during a keynote at SMARTCLOUD SHOW 2026 in Seoul. The report said his lecture covered “hyperscale bug finding,” including work connected to DARPA’s AI Cyber Challenge and Microsoft’s MDASH vulnerability-detection system. ChosunBiz presented Kim’s remarks as the basis for the article; the source does not provide an independent transcript, technical paper, or outside assessment of the systems described.
The report said Kim led Team Atlanta, the winner of last year’s AI Cyber Challenge, a DARPA competition focused on using AI to analyze, detect, and fix software vulnerabilities. ChosunBiz reported that Team Atlanta scored 392.76 points in the final, more than 170 points ahead of the runner-up. In the account, the team’s cyber reasoning system analyzed 28 open-source projects, found 54 of 70 vulnerabilities placed by organizers, fixed most vulnerabilities within an hour, and discovered 18 previously unknown zero-day vulnerabilities. ChosunBiz also reported that more than 90% of the discovered vulnerabilities confirmed as real, although the article does not explain the evaluation methodology in detail.
ChosunBiz reported that the system cost about $152 per successful task, citing Kim, who compared that figure with the much higher cost of human code audits or penetration tests. The report said Team Atlanta included researchers from Georgia Tech, KAIST, POSTECH, and other institutions, with 80% to 90% of the team described as Korean. These figures and comparisons are reported claims from Kim and ChosunBiz, not independently verified measurements in the supplied material.
The report said Kim described limitations in using one AI system to analyze very large codebases. In response, the team used a multi-agent architecture assigning different roles to specialized agents, including vulnerability hunting, static analysis, verification, and patch generation. ChosunBiz said Microsoft’s MDASH extends this idea by first building a threat model from software architecture and past vulnerabilities, then deploying more than 100 sub-agents. Agents acting as hackers, developers, and defenders reportedly check findings against one another, generate proof-of-concept code to test exploitability, and produce remediation patches.
ChosunBiz reported that Kim said MDASH had identified high-risk issues that could potentially allow an external attacker to obtain the highest system privileges by sending packets. Kim also said MDASH performed comparably to newer frontier models despite combining models two generations older, and that it was being used beyond benchmarks in real software-development environments. The supplied report does not identify those environments, disclose the models or test sets involved, provide examples of the vulnerabilities, or include independent confirmation of the claimed performance.
Read the primary source: biz.chosun.com ↗
Why it matters
If independently validated, the approach could lower the cost of security audits and help defenders examine software more quickly. It could also increase the volume of vulnerabilities discovered, creating pressure for software maintainers to verify, prioritize, and fix findings faster.
The practical significance is the possibility of moving parts of vulnerability research from a labor-intensive, episodic process toward continuous automated analysis. ChosunBiz reported Kim’s estimate that AI-based work could reduce the expense of successful security tasks dramatically compared with human audits and penetration tests. Lower costs could make deeper review more accessible to organizations that cannot routinely hire specialist security teams, although the source does not establish that the reported price comparison applies across different software types or operating conditions.
A multi-agent design could address a known operational problem: a general-purpose model may reason well on one task but perform inconsistently on another. In ChosunBiz’s account, separate agents are assigned narrower responsibilities and then asked to verify one another’s findings. That structure may make errors easier to detect, but it also adds coordination, monitoring, and evaluation requirements. The source does not provide enough information to determine whether the additional agents improve accuracy consistently or simply increase compute and operational complexity.
The defensive value could be substantial if systems reliably identify vulnerabilities before attackers do. Faster discovery and patch generation could reduce the time that flaws remain exploitable, particularly in widely used open-source components. At the same time, the same capabilities may be useful to attackers. The article’s description of proof-of-concept generation and packet-triggered privilege-escalation findings illustrates why organizations would need strict controls over access, testing environments, disclosure, and the handling of exploit details.
The report also points to a broader shift in cybersecurity evaluation. Kim said AI models had surpassed human hackers in a March analysis, according to ChosunBiz, and argued that better AI could both make software safer and reveal more weaknesses. That is an important distinction: finding more vulnerabilities does not by itself mean security has improved. The public impact depends on whether maintainers can reproduce findings, prioritize genuine risks, deploy safe patches, and prevent newly generated exploit knowledge from being misused.
The evidence remains limited in the supplied source. ChosunBiz is a reported-secondary outlet, and the article itself says it was translated by AI. No primary MDASH documentation, competition data release, code, benchmark protocol, or independent researcher response is included. The reported results therefore merit attention as an account of Microsoft’s and Kim’s claims, but they should not yet be treated as a general proof that multi-agent systems outperform human security teams or single-model systems in real-world conditions.
What to watch next
The key questions are whether MDASH’s results hold outside benchmarks, how often it misses real vulnerabilities or produces false positives, and how safely its proof-of-concept and patch-generation capabilities are controlled. Public technical documentation and independent testing would clarify the claims.
The first priority is independent replication. Security researchers and software maintainers would need access to clear evaluation procedures, representative codebases, vulnerability labels, and definitions of successful detection, confirmation, and repair. Particular attention should go to the reported 54-of-70 discovery result, the 18 zero-day findings, the more-than-90% confirmation rate, and the $152-per-task estimate. Without methodological detail, those numbers cannot be compared reliably with human audits, automated scanners, or other AI systems.
The next issue is error quality rather than raw discovery volume. Organizations should examine false positives, missed vulnerabilities, duplicate findings, incorrect severity ratings, and patches that compile but change intended behavior or create new flaws. ChosunBiz reported that agents with different perspectives reach consensus, but the source does not show how consensus is measured or whether it correlates with real-world correctness. Independent testing should assess both secure patching and the consequences of an incorrect automated fix.
Deployment safeguards will also matter. A system that generates proof-of-concept exploit code needs isolation, authorization controls, audit logs, and clear rules for responsible disclosure. The source reports that MDASH is being used in real software-development environments, but does not specify the environments, user permissions, human review procedures, or whether generated patches are automatically applied. Those details would determine whether the system reduces risk or introduces new pathways for misuse.
Researchers should compare the multi-agent approach with simpler alternatives, including a single strong model, conventional static-analysis tools, human-led penetration testing, and combinations of these methods. ChosunBiz reported that MDASH achieved performance comparable to newer frontier models while using older models, but did not provide the benchmark design or explain whether cost, latency, coverage, or reliability was the comparison’s main measure. Those distinctions are important for procurement and security planning.
Finally, maintainers and policymakers should watch whether AI-assisted vulnerability discovery changes disclosure workloads. More findings could improve software security if they are accurate and responsibly reported, but could also overwhelm open-source projects and smaller vendors. Meaningful unknowns include the system’s coverage across programming languages and application types, its performance on previously unseen code, the frequency of dangerous patch errors, and the extent of human oversight required before any result is acted upon.
