Microsoft AISpecialist modelLimited access
MAI-Cyber-1-Flash
n/ano search or usage data
Summary
MAI-Cyber-1-Flash is Microsoft’s specialist model for finding and fixing software vulnerabilities. It is deeply integrated into the MDASH harness and designed for large codebases with many security tasks. No public model endpoint or open weights are documented.
Access remains restricted to MDASH and selected customers. Microsoft reports CyberGym results for the combined MDASH system with MAI-Cyber-1-Flash, GPT-5.4, and additional agents. That combined figure is not an isolated MAI-Cyber benchmark.
On this page
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 7 published benchmark results
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
CVEBench (Show measurement, test conditions, and source)
- Source value
- 0.314
- Metric
- published benchmark score; scale not defined in model card
- Unit
- No unit provided
- Benchmark version
- model-card external benchmark
- Category
- cybersecurity
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Microsoft MAI-Cyber-1-Flash model card
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Vendor-reported result. Release checkpoint with RAI safety training; single model in lightweight terminal harness; exact card row CVEBench.
CyberSecEval4 Threat Intelligence (Show measurement, test conditions, and source)
- Source value
- 0.553
- Metric
- published score; scale not defined in model card
- Unit
- No unit provided
- Benchmark version
- model-card external benchmark
- Category
- cybersecurity
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Microsoft MAI-Cyber-1-Flash model card
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Vendor-reported result. Release checkpoint with RAI safety training; single model in lightweight terminal harness; exact card row CyberSecEval4 Threat Intelligence.
CyberSecEval4 Malware Analysis (Show measurement, test conditions, and source)
- Source value
- 0.33
- Metric
- published score; scale not defined in model card
- Unit
- No unit provided
- Benchmark version
- model-card external benchmark
- Category
- cybersecurity
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Microsoft MAI-Cyber-1-Flash model card
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Vendor-reported result. Release checkpoint with RAI safety training; single model in lightweight terminal harness; exact card row CyberSecEval4 Malware Analysis.
CRSBench (Show measurement, test conditions, and source)
- Source value
- 0.651
- Metric
- POV=1200; published score scale not defined in model card
- Unit
- No unit provided
- Benchmark version
- model-card external benchmark
- Category
- cybersecurity
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Microsoft MAI-Cyber-1-Flash model card
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Vendor-reported result. Release checkpoint with RAI safety training; single model in lightweight terminal harness; exact card row CRSBench, POV=1200.
ExploitGym Kernel (Show measurement, test conditions, and source)
- Source value
- 0
- Metric
- published benchmark score; capability measure, not defensive quality
- Unit
- No unit provided
- Benchmark version
- model-card external benchmark
- Category
- cybersecurity
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Microsoft MAI-Cyber-1-Flash model card
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Vendor-reported result. Release checkpoint with RAI safety training; single model in lightweight terminal harness; exact card row ExploitGym Kernel.
- Context
- ExploitGym measures offensive exploit-generation capability; zero is not a general cyber-quality score and should not be ranked against defensive metrics.
ExploitGym Userspace (Show measurement, test conditions, and source)
- Source value
- 0
- Metric
- published benchmark score; capability measure, not defensive quality
- Unit
- No unit provided
- Benchmark version
- model-card external benchmark
- Category
- cybersecurity
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Microsoft MAI-Cyber-1-Flash model card
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Vendor-reported result. Release checkpoint with RAI safety training; single model in lightweight terminal harness; exact card row ExploitGym Userspace.
- Context
- ExploitGym measures offensive exploit-generation capability; zero is not a general cyber-quality score and should not be ranked against defensive metrics.
ExploitGym Browser (Show measurement, test conditions, and source)
- Source value
- 0
- Metric
- published benchmark score; capability measure, not defensive quality
- Unit
- No unit provided
- Benchmark version
- model-card external benchmark
- Category
- cybersecurity
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Microsoft MAI-Cyber-1-Flash model card
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Vendor-reported result. Release checkpoint with RAI safety training; single model in lightweight terminal harness; exact card row ExploitGym Browser.
- Context
- ExploitGym measures offensive exploit-generation capability; zero is not a general cyber-quality score and should not be ranked against defensive metrics.
Similar models
More models of the same kind as MAI-Cyber-1-Flash, chosen by model line, developer, size, price and release date.
- GPT-5.6 CyberOpenAI · released around the same time
- Gemini 3.8 Flash CyberGoogle DeepMind · released around the same time
- GWM Worlds 2Runway · released around the same time
- GPT-Rosalind-5.5OpenAI · released around the same time
- EmbeddingGemma 2Google DeepMind · released around the same time
- Unlimited OCR 2Baidu · released around the same time
- TeleOCRTeleAI · released around the same time
- Kumo-TabularNVIDIA · released around the same time
All specifications
Architecture
Watermarking and provenance
AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.
- AI watermarking
- Unknown
- no source
- Watermarking technique
- Not known
- no source
Licence and access
- Shutdown
- None announced
- no source
Further details
Sources and data date
Benchmark sources at a glance
Which independent evaluators tested this model. This is coverage, not a quality ranking.
No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.
Without vendor claims and historical or disputed results; those stay labelled in the single values.
| Type | Evidence and data date |
|---|---|
| Research status | Research date October 4, 2026. 2 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Unresolved | cybergym-standalone-result Note The 95.95% CyberGym result published in the MDASH article evaluates a composite system with MAI-Cyber-1-Flash, GPT-5.4 and agent tooling, not this model alone; it is not assigned here. |
| Not published | model-card-population-and-scale Note The model card gives no row-level sample counts and does not define the numeric scales for these values. Raw scores are preserved without converting them to percentages. |
| Additional source | Microsoft AI MAI-Cyber announcement (retrieved October 3, 2026; October 15, 2026; October 16, 2026) · Microsoft AI MAI-Cyber announcement · Editorial description reviewed October 3, 2026 |
| Additional source | Microsoft MAI-Cyber-1-Flash model card (retrieved October 16, 2026) |