{"id":107455,"date":"2026-07-28T19:49:37","date_gmt":"2026-07-28T19:49:37","guid":{"rendered":"https:\/\/youzum.net\/microsoft-ai-releases-mai-cyber-1-flash-a-5b-active-parameter-cyber-model-that-pushes-mdash-to-95-95-on-cybergym\/"},"modified":"2026-07-28T19:49:37","modified_gmt":"2026-07-28T19:49:37","slug":"microsoft-ai-releases-mai-cyber-1-flash-a-5b-active-parameter-cyber-model-that-pushes-mdash-to-95-95-on-cybergym","status":"publish","type":"post","link":"https:\/\/youzum.net\/ja\/microsoft-ai-releases-mai-cyber-1-flash-a-5b-active-parameter-cyber-model-that-pushes-mdash-to-95-95-on-cybergym\/","title":{"rendered":"Microsoft AI Releases MAI-Cyber-1-Flash: A 5B-Active-Parameter Cyber Model That Pushes MDASH to 95.95% on CyberGym"},"content":{"rendered":"

Microsoft AI has released MAI-Cyber-1-Flash<\/a>, its first model built specifically for cyber defense. The model does not ship as a standalone endpoint. It runs inside MDASH<\/strong>, Microsoft\u2019s multi-model agentic scanning harness. <\/p>\n

MAI-Cyber-1-Flash<\/strong><\/a><\/h2>\n

MAI-Cyber-1-Flash<\/a> is a transformer with self-attention and sparse Mixture-of-Experts layers. It carries 137B total parameters with 5B active<\/strong>, and a 256k context length<\/strong>. Inputs and outputs are text only.<\/p>\n

It is a cybersecurity-specialized fine-tune of MAI-Code-1-Flash<\/a>, the lightweight agentic coding model already embedded in GitHub Copilot and VS Code. The release describes it as derived from the MAI-Thinking-1<\/a> lineage. <\/p>\n

\n<\/div>\n

\n

Benchmarks<\/strong><\/h2>\n<\/p>

CyberGym<\/a> is a public suite of 1,507 real-world vulnerability reproduction tasks drawn from 188 OSS-Fuzz projects. Microsoft evaluated at CyberGym\u2019s default level 1 configuration, which supplies vulnerable source and a high-level description.<\/p>\n

MDASH running MAI-Cyber-1-Flash alongside GPT-5.4 scores 95.95%<\/strong>. Microsoft frames this as roughly 12 points above Anthropic\u2019s Mythos, and the launch chart places the four competing systems between 83.2% and 85.6%.<\/p>\n

When Microsoft first detailed MDASH<\/a> in May 2026, the harness scored 88.45%<\/strong> on CyberGym using only generally available models. That was already the top public leaderboard score, about five points ahead of the next entry at 83.1%. The research team states the improvement plainly: replacing 80% of the existing models in MDASH moved the harness from 88.4% to 95.95%.<\/p>\n

Why the routing is the real product<\/strong><\/h2>\n

MDASH manages over 100 specialized agents<\/strong> through five stages: Prepare, Scan, Validate, Dedupe, and Prove. Auditor agents flag findings, debater agents argue exploitability (using disagreement as signal), and the Prove stage executes triggering inputs with ASan for C\/C++ targets.<\/p>\n

To control frontier model costs at scale, MAI-Cyber-1-Flash handles up to 90% of MDASH tasks<\/strong>, escalating the hardest 10% to GPT-5.4. This routing yields a 50% cost saving<\/strong> over the previous configuration of GPT-5.4, 5.4 mini, and 5.3 codex.<\/p>\n

MDASH was developed by Microsoft\u2019s Autonomous Code Security (ACS)<\/strong> team, featuring members from the DARPA AI Cyber Challenge-winning Team Atlanta. In May, MDASH-assisted work generated 16 CVEs<\/strong> (including four Critical remote code execution flaws) in the Windows networking and authentication stack. Retrospectively, it recovered 96% of 28 MSRC cases in <\/strong>clfs.sys<\/strong> and 100% of 7 cases in <\/strong>tcpip.sys<\/strong> over a five-year window.<\/p>\n

Performance<\/strong><\/h2>\n

The research team present standalone results from a lightweight terminal harness:<\/p>\n

\n\n\n\n\n\n\n\n\n\n
Benchmark<\/th>\nMAI-Cyber-1-Flash<\/th>\n<\/tr>\n<\/thead>\n
CVEBench<\/td>\n0.314<\/td>\n<\/tr>\n
CyberSecEval4 \u2014 Threat Intel<\/td>\n0.553<\/td>\n<\/tr>\n
CyberSecEval4 \u2014 Malware Analysis<\/td>\n0.33<\/td>\n<\/tr>\n
CRSBench<\/td>\n0.651 (POV=1200)<\/td>\n<\/tr>\n
ExploitGym \u2014 Kernel \/ Userspace \/ Browser<\/td>\n0 \/ 0 \/ 0<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

The straight zeros on ExploitGym are deliberate, not a defect. Microsoft team states the model was trained to perform defensive tasks such as patching bugs, and not offensive tasks such as deploying malware. A 5B-active model that cannot generate exploits but can drive a 95.95% discovery pipeline is exactly the artifact a defender-only product needs.<\/p>\n

How to use it<\/strong><\/h2>\n
\n<\/div>\n

\n

Key Takeaways<\/strong><\/h2>\n