Microsoft announced MAI Cyber 2 Flash on November 17, 2026, introducing a specialist AI model for vulnerability discovery, validation, remediation guidance and security investigation. The model has a one-million-token context window and will be delivered through Microsoft’s security products, with access limited to eligible customers rather than a general-purpose model API. Microsoft’s announcement also sets an operational deadline: automatic replacement of the existing MAI-profile model in Codename MDASH begins in mid-December.
Microsoft reports strong benchmark results and a substantial cost advantage over larger models. The most consequential detail, though, is its benchmark-contamination disclosure. The company excluded 92 of 905 CyberGym-E2E tasks after identifying overlap between their underlying projects and vulnerability reports and its training data.
That leaves two practical questions for security teams: what the published numbers actually measure, and what changes when Microsoft updates the model inside an existing security workflow.
A Million-Token Model Inside a Larger Security System
MAI Cyber 2 Flash is built to work inside Microsoft’s security tooling. Its model page describes the deployment boundary directly: “Built exclusively for Microsoft security.”
The one-million-token context window gives the model capacity to work with a large amount of information during an investigation. Tokens are the units an LLM processes, including fragments of words, code and punctuation. They are not equivalent to lines of code or files, so the advertised window cannot be translated into a fixed repository size.
For vulnerability work, the useful implication is that more related material can remain available together: source files, call relationships, patch history and investigation notes. A flaw may depend on how several components interact, making context capacity relevant beyond simply accepting a larger prompt. The window size alone says little about whether the model will identify the right dependency or produce a correct fix.
Microsoft places the model inside Codename MDASH, its multi-agent security scanning system. The published workflow includes:
- Preparation: Build a per-repository knowledge base with a threat model, repository map, call graph and CVE or patch information.
- Discovery: Use more than 100 specialized AI agents to scan for vulnerabilities.
- Validation and deduplication: Cross-check findings with multiple models, filter false positives and collapse duplicates.
- Proof and remediation: Generate a proof of concept, suggest a patch and check the patch’s correctness.
Microsoft says MAI Cyber 2 Flash handles up to 90% of MDASH tasks, reserving larger models for the hardest cases. That is a workload-allocation claim, not a claim that the model finds 90% of vulnerabilities.
My view is that this distinction matters more than the “Flash” branding. Customers receive the output of an orchestrated system, where model selection, tools, cross-checking and patch validation all affect the result. A standalone model score captures only part of that workflow.
What Microsoft’s Benchmark Results Show
Microsoft reports results on CyberGym-E2E, a vulnerability-discovery benchmark, and WebExploitBench. These are vendor-run evaluations; the supplied coverage provides no independent reproduction.
The comparison conditions differ, as NeoTeo’s account of the announcement explains:
| Benchmark | MAI Cyber 2 Flash | Comparison model | Reported conditions |
|---|---|---|---|
| CyberGym-E2E | 92.3% task success | GPT-6 Astra: 91.5% | MAI used half the reasoning budget; all models scored on the same 813-task subset |
| WebExploitBench | 40.4% | GPT-5.6 Cyber: 38.4% | Same reasoning budget |
The CyberGym-E2E lead is 0.8 percentage points. The WebExploitBench lead is 2 percentage points. Neither should be described as a general accuracy advantage across security work: they measure different tasks and use different comparison models.
A reasoning budget describes the allowance given to a model’s reasoning process during evaluation. Half the budget is not automatically half the elapsed time or half the total system cost, particularly when tools and multiple agents are involved.
There is also a relevant comparison missing. According to Aivio’s coverage of Microsoft’s disclosure, Microsoft had not evaluated the model against the dedicated GPT-6 Cyber model, citing restricted access and additional safeguards needed for agentic, tool-using evaluations. The published results therefore support a narrower claim than leadership over every available security model.
Why Removing 92 Tasks Changes the Benchmark Claim
Microsoft says some CyberGym-E2E tasks used the same open-source projects and vulnerability reports found in portions of its training data. It says the benchmark’s reference solutions were not used for training. The company removed the overlapping tasks for every evaluated model, then reported results on the remaining 813.
The arithmetic is straightforward: 92 divided by 905 is approximately 10.2% of the original task set. The reported 92.3% score therefore belongs to a filtered benchmark, not the complete 905-task suite. Both NeoTeo and Aivio describe that exclusion.
Training overlap matters because prior exposure to a project or vulnerability report could make a task easier without demonstrating the same ability to investigate unfamiliar software. Exposure to source material is different from training on the benchmark’s reference answer, but it can still complicate interpretation.
Applying the same exclusions to competitors makes the reported comparison internally consistent: every model faces the same remaining tasks. It also changes the test population. Without the omitted-task results and further analysis, readers cannot determine how removing those tasks affected the models’ relative positions.
I would treat this disclosure as a strength of the documentation, while keeping the claim bounded. Microsoft identifies the overlap, quantifies it and explains its response. That makes the result easier to assess. It does not prove that every possible overlap was detected or tell us how the model performs on an organization’s private code.
The practical lesson is to preserve the denominator. Any internal presentation quoting “92.3% on CyberGym-E2E” should also say “813 tasks after exclusions.”
The Cost Advantage Is an Evaluation Calculation
Microsoft’s headline economic claim is that MAI Cyber 2 Flash costs less than 2% of GPT-6 Astra per task on its CyberGym-E2E evaluation.
The underlying calculation uses token consumption and list prices for input, cached-input and output tokens. Aivio reports these rates from Microsoft’s footnotes, all per one million tokens:
- MAI Cyber 2 Flash: $0.10 input, $0.01 cached input and $0.50 output.
- GPT-6 Astra on Azure OpenAI Global: $10 input, $1 cached input and $50 output.
Frequently Asked Questions
3 questions
1Is MAI Cyber 2 Flash Available Through a Public API?
No. Microsoft is delivering it through its security products, with MDASH access restricted to eligible, verified defenders.