Voting equipment that is connected to the internet
A University of Michigan team took over Washington D.C.'s internet voting pilot in 48 hours — and officials didn't notice for days. Here's what that means for every election run on a networked machine.
It is a Tuesday morning in September 2006, and in a Princeton University computer lab, three researchers are holding a physical voting machine — a real one, an Diebold AccuVote-TS, the kind deployed across American precincts. They have had it for less than a day. Within roughly sixty seconds of access, they have installed software that will steal votes, cover its tracks, and spread itself automatically to every other machine it touches during normal election operations.
No network connection required. A memory card and a minute of physical access was enough.
Now add a network. The story gets considerably worse.
The machine in the room you can't see
Every voting machine is a computer. And every computer that talks to another computer — through a modem, a cellular connection, a Wi-Fi radio, an ethernet port, or even a "poll book" sync cable — is, in security terms, a networked computer. The question isn't whether internet-connected voting equipment is a theoretical risk. It's whether the officials running your election can tell the difference between a machine that is isolated and one that isn't — and whether you have any way to check their answer.
That distinction matters enormously because the whole security argument for electronic voting has historically rested on a premise: the machines are air-gapped. They don't talk to the internet. They are islands.
When researchers poke at that premise — in a lab, in a live pilot, with a university team and a 48-hour window — the islands start to look like peninsulas.
Washington, D.C., September 2010: 48 hours
The District of Columbia Board of Elections did something that deserves credit: before deploying an internet ballot-return system for overseas and military voters, it ran an open public trial and explicitly invited the public to probe it.
A team led by University of Michigan computer scientist J. Alex Halderman — the same researcher whose name appears on the Princeton voting-machine analysis four years earlier — accepted the invitation. Within roughly 48 hours, the team had gained near-complete control of the pilot server by exploiting a software vulnerability. They could change every ballot that had been cast. They could read ballots that were supposed to be secret. And — as a demonstration that the intrusion was real and they were in complete control — they programmed the system to play the University of Michigan fight song after each simulated vote.
Officials did not detect the intrusion. The researchers had to disclose it themselves.
The D.C. Board of Elections cancelled plans to use the system for live binding votes. The findings were later published as "Attacking the Washington, D.C. Internet Voting System" in the peer-reviewed Financial Cryptography conference proceedings.
Here is the part worth sitting with: the attack worked because the team disclosed it. Had they been a hostile actor rather than a research team flagging vulnerabilities for the public good, the ballots would have been changed and no one would have known.
The D.C. pilot is now a canonical example in election-security research. But it did not change the underlying condition it exposed: the assumption that internet connectivity is absent, controlled, or containable in systems that handle votes.
Read the researchers' paper here.
"Air-gapped" is a claim, not a guarantee
Here is what officials mean when they say voting machines are not connected to the internet: the machines are not supposed to be. They were designed to be isolated. They are certified to a standard that did not permit live connectivity.
Here is what that claim does not cover: the electronic poll books that sync voter rolls — often via cellular modem. The election management systems that receive results from tabulators over local networks that may share infrastructure. The firmware updates delivered over removable media that passed through a connected computer somewhere up the chain. The Wi-Fi radio that was left enabled in a device because disabling it was an extra configuration step someone skipped at 2 a.m. during setup.
A machine that is "not connected to the internet" and a machine that is verifiably, independently confirmed to be isolated are not the same thing. The first is a policy. The second is an auditable fact. Most election infrastructure runs on the first and claims credit for the second.
The 2005 GAO report on electronic voting (GAO-05-956) flagged this directly: federal efforts to improve electronic voting security and reliability were underway but incomplete, and key activities — including sharing information about system vulnerabilities — had not been finished. That report was written eight years before the D.C. pilot. The gap it identified was structural, not incidental.
Details of the GAO findings here.
The Princeton paper: what happens on a connected machine
Go back to the Princeton lab in 2006. The Feldman, Halderman, and Felten paper demonstrated something that the D.C. pilot would later confirm in a live environment: that a compromised voting machine does not need to announce itself. The malicious code they built altered all internal records, logs, and counters to stay internally consistent. The machine would pass a standard diagnostic check. A routine audit, looking for discrepancies in the machine's own logs, would find none.
They also built a virus — in the literal computer-science sense. Code that spreads from machine to machine during normal election-night procedures, the same memory-card exchanges that happen in every county, in every precinct, on every election night.
Now consider what changes if the machine has any connectivity at all. A networked compromise does not require physical access. It does not require a rogue insider with a memory card and sixty seconds alone with a machine. It requires a software vulnerability on a device that talks to something else — a poll book, a results-transmission system, a remote-access tool left active for technical support.
The Princeton paper is not primarily about what a one-minute physical attacker can do. It is about what a connected system with closed, unverifiable software looks like from the inside: a machine that can be wrong in every way that counts while appearing correct in every way you can check.
The connectivity problem in the wild
The D.C. pilot is the clearest documented case of a live internet-based voting system being compromised in a public test. But the underlying issue — connectivity assumed absent, found present — is not unique to online ballot return.
California's 2007 top-to-bottom review, in which University of California computer scientists examined the source code of the state's certified voting systems, found that the Diebold system was vulnerable to malicious software and to viruses spreading between machines. The Hart InterCivic system had unsecured network interfaces. These were systems not intended to be connected to the internet. The vulnerabilities existed anyway, waiting for a pathway.
Ohio's EVEREST review the same year reached a similar conclusion across all three systems it studied: critical security failures that render their technical controls insufficient to guarantee a trustworthy election. Among the shared failures: insufficient defense against insider attacks, improperly implemented cryptography, and audit logs that could be altered or erased by the people they were meant to monitor.
If the logs can be faked, "we checked the logs and found nothing" is not evidence of safety. It is evidence that the check was run.
The New South Wales iVote system in 2015 — a live online election with roughly 280,000 votes already cast — was found by independent researchers Halderman and Teague to load code from an external analytics server over a connection vulnerable to known TLS weaknesses. An attacker positioned on the network could potentially have manipulated votes and compromised ballot privacy. The election authority's pre-election security review had not flagged this. The researchers published their findings while voting was still underway.
The pattern across these cases is not incompetence. It is a structural feature of complex software systems: every connection point is an attack surface, and in systems that are not built for public verifiability, there is no independent way to confirm that all of them are closed.
What "certified" actually certifies
When a voting system receives federal certification from the U.S. Election Assistance Commission, it has passed a set of tests run by an accredited lab at a point in time. The tests check whether the system meets the Voluntary Voting System Guidelines as they were written when the system was submitted.
Certification does not confirm: whether the specific unit deployed in your precinct is running the same software that was tested. Whether a modem, Wi-Fi radio, or remote-access tool was added or left active after certification. Whether the network it touches has been compromised. Whether the firmware on a component was updated after the certification date.
"Certified" means it passed a test once. It says nothing about the machine in the room on Election Day.
This is not a theoretical concern. The Coffee County, Georgia episode in 2021 — where an entire jurisdiction's voting system software was copied by people let in by county officials — involved software that was certified and in active use. Certification did not prevent the copy. It did not detect it. It did not create any mechanism by which the public could verify what had happened to the code they were voting on. That episode is documented in federal court depositions in the Curling v. Raffensperger case.
The verifiability gap the officials can't close for you
Here is the core problem with internet-connected — or plausibly connected — voting equipment: connectivity creates a class of attacks that leave no evidence the attacked system can self-report.
Paper ballots, marked by voters and counted by humans or optical scanners, have an independent physical existence. A machine tally can be checked against them. This is why Georgia's 2020 statewide hand count of roughly five million ballots — triggered by the narrow margin — could meaningfully confirm the machine result. The paper was there. It was checkable. The variation was about a tenth of one percent.
An internet-compromised voting machine, or a compromised results-transmission system, may have no equivalent independent check if the system was designed to be the authoritative record. When the D.C. team changed ballots in 2010, those changes would have been the ballots of record. The paper analogy is not "someone marked the wrong number on a tally sheet." It is "someone replaced the ballots themselves, and the new ballots are indistinguishable from the originals."
The German Constitutional Court articulated the standard that matters here, in its 2009 ruling striking down voting computers: the essential steps of voting and of determining the result must be examinable by a citizen without specialist knowledge. A system whose correctness requires trusting closed software, opaque network paths, and unverifiable logs fails that standard regardless of whether any tampering ever occurred.
The November 2020 CISA joint statement — which declared the election "the most secure in American history" — grounded its own confidence in the existence of paper records that allow recounts and audits. That is the statement's own reasoning. It is an implicit acknowledgment that security flows from verifiability, not from official assertion.
What would actually close the gap
The D.C. team was detected because they made noise — a fight song playing on a government server is not subtle. A sophisticated attacker would not play the fight song.
The fix is not better intrusion detection. Intrusion detection is a layer that helps, but it is a trust layer: you are trusting the detection system, the people who monitor it, and the logs it generates. The fix is building election systems where the correctness of the count is independently verifiable by anyone, without having to trust any of those layers.
That means:
- Voter-marked paper ballots as the authoritative record — not QR codes the voter cannot read, as the federal court found in Curling v. Raffensperger.
- Results published at the precinct level, in machine-readable formats, the moment each precinct finalizes — not aggregated hours or days later in ways that hide the underlying data.
- Cryptographic, end-to-end verifiable tallying where a voter can confirm their encrypted ballot is included in the count without revealing how they voted — the standard Switzerland's 2019 public code review showed is achievable but must be implemented without trapdoors. That review found one.
- Public adversarial testing before deployment, not after — the approach Brazil's Superior Electoral Court institutionalized with its annual public security tests. The D.C. pilot worked as a safety valve because it was tested before votes were real. Certification after a closed lab test is not the same thing.
The internet-connectivity gap is not primarily about whether a specific machine has a modem. It is about whether the design of the counting system lets anyone — voter, researcher, opposing candidate, court — independently confirm the result without taking the word of the officials or the vendor.
A result that requires you to trust the network is a result that requires you to trust everyone on the network.
What you still cannot verify — and what would change that
Right now, in most U.S. jurisdictions and in most democracies that use electronic voting, there is no public mechanism to confirm:
- Whether the machine you voted on was running the software it was certified to run.
- Whether the results-transmission system between precinct and central tabulator was accessed during the count.
- Whether the election management system that programs the machines was itself connected to anything external in the weeks before the election.
Officials may be able to tell you those things. Vendors may have logs. Auditors may have checked. Those are all claims — made by parties with interests and limited independent accountability.
What would change it: precinct-level, machine-readable results published in real time, alongside cryptographic ballot receipts voters can check themselves, and open-source software whose behavior can be independently confirmed to match what is running on Election Day. Not a promise that the right software is installed. A proof.
That is the standard the D.C. 2010 pilot failed to meet — and would have failed to meet even without the Michigan team, because there was no public mechanism to confirm it was secure. The fight song was not the breach. It was just the evidence.
See how widespread the internet-connected equipment gap is across election systems globally. →
Explore the full map of election integrity gaps. →
Read the two-minute version. →
Sources
- Wolchok, Wustrow, Halderman, Prasad — Attacking the Washington, D.C. Internet Voting System, Financial Cryptography 2012
- Feldman, Halderman & Felten — Security Analysis of the Diebold AccuVote-TS Voting Machine (USENIX/EVT)
- GAO-05-956, 'Elections: Federal Efforts to Improve Security and Reliability of Electronic Voting Systems Are Under Way, but Key Activities Need to Be Completed' (Sept. 21, 2005)
- Halderman, Teague — The New South Wales iVote System: Security Failures and Verification Flaws in a Live Online Election, E-Vote-ID 2015 (arXiv:1504.05646)
- California Secretary of State, Top-to-Bottom Review (2007) — decisions and the independent experts' reports
- Source Code Review of the Diebold Voting System (University of California, 2007)
- Source Code Review of the Hart InterCivic Voting System (University of California, 2007)
- EVEREST: Evaluation and Validation of Election-Related Equipment, Standards and Testing — Final Report (December 7, 2007)
- Lawfare — What the Heck Happened in Coffee County, Georgia?
- Curling v. Raffensperger, No. 1:17-cv-2989-AT, Opinion and Order (N.D. Ga. Oct. 11, 2020) (Justia)
- Bundesverfassungsgericht, Judgment of 3 March 2009, 2 BvC 3/07 and 2 BvC 4/07 (English translation)
- Joint Statement, Election Infrastructure Government Coordinating Council & Sector Coordinating Council (Nov. 12, 2020) (CISA)
- Lewis, Pereira, Teague — Ceci n'est pas une preuve (trapdoor commitments in the Scytl-SwissPost Internet voting system), 2019