MAI-Cyber-1-Flash
Microsoft · Veröffentlicht Juli 2026
Microsoft's first in-house cybersecurity model, built on the MAI-Thinking-1 family and embedded inside the MDASH multi-agent harness, handles about 90% of security workloads at roughly half the cost of prior configurations and reserves GPT-5.4 for the hardest 10%. Microsoft reports 96% on CyberGym by the any-crash metric and 86.3% by the stricter final-submission metric, both from a tuned full-system configuration rather than an independent model-versus-model test. Available only through Azure AI Foundry with vetted enterprise access; no standalone pricing or public API.
Ist es das Richtige für dich?
Gut für
- Vulnerability discovery on CyberGym — 96% on the any-crash metric and 86.3% on the stricter final-submission metric, both Microsoft-run on the MDASH full system
- Cost efficiency at scale — specialized small model handles about 90% of tasks, reserving GPT-5.4 for the hardest 10%, at roughly half the cost of prior MDASH configurations
- Deep integration with Microsoft's security estate — trained on decades of enterprise defense telemetry
Nicht geeignet für
- Not independently benchmarked — all CyberGym scores are vendor-reported from a full-system configuration (harness plus dual models), not a controlled model comparison, and rival scores may use a different CyberGym metric
- Restricted access — no public API or self-serve tier; gated behind Azure AI Foundry vetting
- Dual-use risk — built to find hard vulnerabilities, the same capability serves attackers with no built-in gating beyond access controls
Leistung nach Aufgabe
Vulnerability discovery
96% any-crash and 86.3% final-submission on CyberGym, Microsoft-run on a full-system configuration; not comparable like-for-like with other vendors' scores
Code patching and remediation
Generates working patches and validates them — green-team loop closes the fix cycle
General coding
Purpose-built for security; underperforms general-purpose models on non-security coding tasks
Reasoning
Competent within security domain but defers 10% of hardest queries to GPT-5.4
Preise
Eingabe
N/A (MDASH SCU consumption)
Ausgabe
N/A (MDASH SCU consumption)
Kontext
Azure AI Foundry vetted access
Benchmarks
Urteilsverlauf
Quellen
Prüfprotokoll
- Status· Keine Änderungen
Automatisierter Agent
- Inhalt· Aktualisiert
Automatisierter Agent
CyberGym figures conflated: Microsoft reports 96% any-crash, 90.4% target and 86.3% final-submission. Summary (95.95%, 12 points over Mythos 5) corrected via verdict_set (conditional held); strengths, task_strengths and benchmarks written with the metric named. Cross-vendor lead dropped as not like-for-like.
- Profil· Keine Änderungen
Automatisierter Agent
The launch post still gives MDASH-embedded use, 50% of the cost of leading models, and GPT-5.4 for the hardest 10%, with no standalone pricing. Flag for a follow-up, not written: a summarised read says the 96% CyberGym figure is an 'any-crash' score and a separate leaderboard 'final-submission' score is 86.3%, while the stored 95.95% strength compares it head-to-head with Mythos 5. This needs a verbatim read before any benchmark text changes.
- Preise· Keine Änderungen
Automatisierter Agent