A Voting System Vulnerability: AI Reconstructed the Order of the Ballots

Bruce Schneier has described how an old vulnerability in a voting system took on new life thanks to artificial intelligence tools. The flaw in question makes it possible to reconstruct the order in which ballots were cast. It was first reported nearly four years ago, yet the problem was never fixed: the scanners it affects are still in use in 21 U.S. states.

Schneier demonstrated that the method works during the May 2026 primary in Georgia — one of those states. The result was an analysis of voter behavior obtained, in his own words, without a single questionable act. He didn't touch the voting machines, didn't break into a network, didn't study source code, and didn't access anything restricted.

Public data only

The entire “hack” came down to Schneier pointing a coding AI agent at the original research paper describing the vulnerability, and then handing it the two datasets mentioned in that paper. The first was the list of voters who voted early, by precinct. The second was the CVR file (cast-vote record) — that is, a record of every ballot and the choices made on it. The CVR contains no names or other identifying information.

The paradox is that it is precisely the openness of this data that creates the problem. The CVR file is provided on request entirely by design: a public, ballot-level record is what makes it possible to independently verify election results. Transparency, conceived as a guarantee of an honest count, when combined with the scanners' vulnerability and the publicly available list of early voters, turns into the ability to link specific ballots to voter behavior.

What AI changed

What is telling here is not so much the very existence of a flaw that has been known for years, but the ease of exploiting it. Previously such an analysis required special training and effort. Now it is enough to point an AI agent at a description of the vulnerability and supply it with open data — the agent does the rest itself. A technically complex attack turns into a routine task, available to anyone with access to public files and a modern code-writing tool.

This storyline echoes a broader theme that Schneier raises: AI agents change not the nature of vulnerabilities, but the threshold beyond which they can be exploited. What had remained for years a theoretical problem from an academic paper became a practical tool for analyzing real elections.

Source: Schneier on Security