MAI-Cyber-1-Flash at a glance

MAI-Cyber-1-Flash is Microsoft's first specialist cyber model. It is a compact, code-heavy descendant of the MAI-Thinking-1 lineage, tuned to reason across complex codebases and help identify vulnerabilities that simple pattern matching can miss.

The model is currently delivered inside Codename MDASH, Microsoft Defender's multi-model agentic code scanner. Microsoft does not document a direct MAI-Cyber API, downloadable weights or a consumer playground for this model.

MDASH builds language-aware indices, maps call paths and the attack surface, ranks risky code, runs more than 100 specialist agents, debates candidate findings across models, deduplicates results and, where possible, constructs triggering inputs in a sandbox to help validate a vulnerability.

Microsoft says MAI-Cyber-1-Flash is efficient enough to handle up to 90% of MDASH tasks, with a larger model such as GPT-5.4 used for the hardest 10%. The published comparison describes roughly 50% lower cost than the earlier GPT-5.4, 5.4 mini and 5.3 Codex configuration; that is a relative system result, not a customer price list.

Microsoft reports 95.95%, rounded to 96%, on CyberGym's any-crash score, 90.4% on target any-of and 86.3% on final submission. These measures test candidate exploit-triggering inputs under benchmark conditions; they do not establish recall across a customer's code, an acceptable false-positive rate or freedom from vulnerabilities.

MDASH is read-only. It produces SARIF, an HTML view, severity and confidence signals, and remediation guidance, but does not patch the environment, open pull requests or merge changes. A separate Defender CLI plus GitHub Copilot flow can propose a local function-level fix for developer review.

The current FAQ calls MDASH a private preview and requires an eligible Microsoft Defender XDR plan plus Foundry configuration. Microsoft Learn documents approximately 256 MB per repository and one concurrent scan per tenant during this phase, both subject to change.

Microsoft says source and scan results are persistently stored, encrypted at rest, in the customer's Defender XDR tenant region, while inference runs in the region selected for connected Foundry endpoints. Preview terms under the Microsoft DPA apply, and retention and deletion timelines are handled separately.

This is an additional security-analysis layer, not a replacement for threat modeling, SAST, dependency and secret scanning, dynamic testing, fuzzing, penetration testing, secure review, incident response or an accountable release decision.