# MAI Cyber 2 Flash: Prepare for the MDASH Model Switch

What Microsoft's December MAI Cyber 2 Flash rollout means for MDASH teams, and how to evaluate findings without mistaking benchmark cost for your bill.

By Paras Madan | 2026-11-18


Microsoft announced MAI Cyber 2 Flash on November 17, with an automatic MDASH model replacement starting in mid-December for eligible customers. Teams using the MAI profile should prepare a comparison before the findings landing in their security queue change.

My recommendation is to treat this as a change to your security review workflow, not a reason to expand scanning scope overnight. Establish what the current system catches, what your reviewers reject, and how much work a useful finding creates.

## Check whether the rollout applies to you

Microsoft's November 17 announcement says Cyber 2 Flash is coming exclusively to Microsoft security products, including MDASH. It will automatically replace the existing model in the MAI profile for eligible customers starting in mid-December. That is not an announcement of a generally available standalone API. The model has a one-million-token context window. [Read the launch announcement](https://microsoft.ai/news/introducing-mai-cyber-2-flash/).

If you already use MDASH, ask your account contact which repositories and profiles are affected, when your deployment changes, and what version information you can retain. Ask whether a comparison period or rollback is available rather than assuming either exists.

If you do not have access, confirm eligibility first. Rebuilding your security pipeline around an unavailable endpoint would be a poor use of the next month.

## Keep the benchmark in its lane

Microsoft reports 92.3% on CyberGym-E2E versus 91.5% for GPT-6 Astra, using half the reasoning budget and less than 2% of its offline per-task cost. The evaluation excludes 92 of 905 tasks because their projects and vulnerability reports overlapped training data; every model was scored on the remaining 813. The company explicitly distinguishes this cost calculation from customer spend. [See the methodology notes](https://microsoft.ai/news/introducing-mai-cyber-2-flash/).

I would not put a 98% saving into a budget spreadsheet from that result. Your useful unit is the cost of a confirmed issue your team can fix, including review time and failed investigations.

## Evaluate the workflow, not just the model

Microsoft's [Cyber product page](https://microsoft.ai/models/mai-cyber-2-flash/) describes MDASH as a multi-agent process covering repository preparation, discovery, validation, deduplication, reproduction and patch checks. Larger models handle harder cases.

That matters when comparing two runs. A different harness configuration can change the result even if the underlying model improves. Keep a record of the repository revision, enabled checks, exclusions and review policy alongside model details.

For an initial comparison, I would choose one repository with an owner who can review the output promptly. Include previously fixed issues and representative clean code, keeping evaluation examples separate from any customization material. Do not use a live production incident as your first controlled comparison.

## Measure reviewer work explicitly

My proposed scorecard has four fields: confirmed issues, unsupported findings, duplicate reports and time to a reviewable fix. Record why each finding was accepted or rejected.

A longer report is not automatically a better report. Ask whether it identifies the affected code, explains the relevant conditions and gives the maintainer enough evidence to make a decision. Track cases where the suggested patch removes a symptom but breaks expected behavior.

Keep human review for changes. A system finding a vulnerability and a system producing a safe, compatible patch are different achievements.

## Decide what would justify broader use

Before rollout, write down your acceptance criteria. For example, require no regression on your known-issue set, an acceptable false-positive burden and patch suggestions that pass the project's existing checks.

Those are proposed evaluation gates, not results I have measured with Cyber 2 Flash. If the new profile clears them, expand gradually. If it does not, preserve the evidence and escalate the specific failure rather than debating a leaderboard.

The useful outcome is a more reliable security queue. Cheaper reasoning only helps if the people responsible for the code can act on it.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help developer-tool teams explain model changes through practical technical writing and launch stories.

[Book a 30-minute call](https://cal.com/paras-madan/30min).

