AI Finds Federal Software Flaws—Who Prioritizes the Fixes?

1 What the scanner claims

On September 8, 2026, Microsoft announced that codename MDASH brings agentic AI security scanning to US government customers through Azure Government, as a feature of Microsoft Defender. The company says it works by putting more than 100 specialized AI agents, using a variety of models, on the same body of code, with each trained to recognize a different category of weakness. Findings then go to a second group of agents that argues for and against whether each flaw is genuinely reachable. It is available to agencies now; the performance figures are Microsoft-reported.

2 Scanning is the easy stage

A vulnerability, in the standard definition, is a flaw in a system design, implementation, or management that a malicious actor could exploit to compromise security. Finding candidates is the first stage of a five-stage pipeline, and it is the cheapest to automate. The expensive stages are validation, routing, fixing, and confirming the fix actually closed the exploit path.

3 Who validates a finding

Microsoft describes an adversarial review inside the tool: one group of agents flags a weakness, a second argues whether it is reachable and dangerous, and results are merged, deduplicated, and where possible demonstrated rather than asserted. That is internal validation. It does not replace an agency security team deciding whether a flagged flaw maps to a real system, a reachable path, and a mission consequence.

4 Prioritization is an agency decision

Microsoft says agencies are ranking their software by mission importance and working down the list, and that cost reduction moves the coverage goal closer. Volume of findings is not improvement. The agency has to set the order, because the same flaw carries different weight on a public-facing benefits portal than on an isolated test system.

5 Measuring improvement, not volume

The measures that matter are time from discovery to patch, the share of validated high-severity findings closed within a set window, and whether a patched flaw is confirmed unreachable afterwards. Microsoft frames the durable defender advantage as time, meaning the gap between when a weakness can be found and patched and when someone else finds it. That framing argues for cycle-time metrics over finding counts.

6 The cost curve and coverage

Microsoft says its MAI model family is designed to make scanning affordable at scale and that its newest addition is expected to cut the cost of an individual scan roughly in half. It also notes there is far more code in the world than can be reviewed at this depth. Cheaper scans expand coverage; they do not by themselves expand the number of engineers who can fix what is found.

7 The bottom line

Discovery is automated; prioritization is not.

Validate findings against reachable, mission-critical systems before fixing.

Measure time to verified patch, not the number of alerts.