Pulling Back the Curtain on the Voting Booth
This recording is not on YouTube. It plays from the DEF CON media server; the button opens the file there.
Where
Voting Village
DEF CON 34
Duration
31 min
Shows how a known Dominion voting machine flaw, a predictable record-ID sequence, let an AI agent link Georgia ballots back to individual voters.
Summary
The talk revisits a 2022 flaw in Dominion voting systems. Scanners tag each ballot with a number meant to be random for shuffling the records, but the numbers come from a fixed arithmetic sequence. Because the sequence is predictable, anyone with the public cast vote records can reverse it, recover the order in which ballots were scanned, and tie ballots to voters. The flaw does not change tallies, but it breaks the promise that nobody can tell how a person voted.
The speakers describe how a Georgia advocate with no coding background used a chatbot to reproduce the recovery from public Georgia files and the research paper. The agent reportedly grouped voters into buckets by check-in time, party and ballot type, and claimed it could attribute roughly 1.5 million in-person ballots. The speakers stress that these are the model's own claims, that models make mistakes, and that the work took about an hour or two on a commute, which they argue shows that exploits once requiring specialist weeks are now quick to run.
The session closes with the policy fight in Georgia. The Secretary of State declined to apply the available vendor patch and ordered counties to stop releasing records with record IDs, a step the speakers say does not stop insiders and vendors from accessing them. They ask for the patch, redacted records, sanitizing tools for old data, and no new records containing reversible numbers. Their broader point is that obfuscation is not a substitute for real randomness or encryption, and that AI lowers the barrier to exploiting known weaknesses.
Materials
Related talks
Policy Village·33 min
Tati
Traces how Chinese, Soviet, German and American surveillance histories shape the privacy defaults that AI products are now built on.
BioHacking Village·14 min
Christine Von Raesfeld
Shows how a rare-disease patient's involvement changed what AI-assisted student teams investigated in a Stanford course built on multi-omic data.
Crypto Privacy Village·27 min
Stephanie Ding · Robin Cheng · Erika Lee
Presents Cove, an open-source framework that chains attested enclave computations into multi-party workflows where private inputs stay hidden and the results can be verified.


