John Odum is the Montpelier city clerk and a certified ethical hacker. I know the implications of the supposedly coming “AI apocalypse” and how it relates to elections are on people’s minds. I’ve heard it. So let’s talk about it. Most of us have probably heard about the Hugging Face incident. During an OpenAI cybersecurity evaluation, AI agents circumvented controls designed to contain them because they were basically struggling with their homework — a cybersecurity challenge assigned to them. They took aim at the Hugging Face platform. Relentlessly seeking success and despite controls, they reached the open internet and penetrated Hugging Face’s systems. Since then, six more instances of its models behaving deceptively or working around safeguards have been discovered.  This incident is important because the AI agents’ objective was mundane: pass a test. When solving the assigned challenge became difficult, they didn’t accept failure. The agents just inferred where the answers might be and pursued them. No human directed this attack — the agents improvised it to complete their task. So could something like this happen to voting systems? The first question we have to ask about the dreaded AI apocalypse we’re all being warned about is “Why?” Why would a self-directing AI mess with an election? Probably not because it cared about the election for its own sake. It would do so because a particular outcome helped it achieve another objective — perhaps one assigned by a human, or maybe one it arrived at while pursuing a broader goal. The election would not be the objective, but a means to an end. Approaching this as a hacker rather than an AI professional, consider how things work at present. Election-management systems, databases and other pieces of networked infrastructure could be targets, and AI doesn’t have to become the full-on existential threat of an artificial sentience crafting its own agenda. Acting on somebody else’s agenda is apocalyptic enough when you’re talking about a technological threat that refuses to lose. Project even further forward to the time of quantum computers capable of breaking many of today’s encryption systems, and the prospect becomes scarier still.  Nobody assigned a social or political agenda, or machine consciousness, to make the Hugging Face incident happen. In the scenario everyone is concerned about, an AI could weaponize itself against free and fair elections. You need only an agent with a political objective, enough autonomy and access to pursue it, a goal — assigned or self-generated — that it keeps pursuing until it succeeds and the ability to improvise. An empowered AI agent could pursue an assigned outcome across the election ecosystem. That means voter-registration databases, electronic pollbooks, election-night reporting, candidate-filing systems, ballot-tracking portals, municipal email, vendor-support accounts, state networks, public-information channels and election-management systems. Voting equipment can be isolated, but by necessity, elections as an administrative process are not. READ MORE Simple, high-profile attacks could include erasing datasets and crashing election-management systems. Those would be high-profile and obvious, designed to crush confidence in the system — again, should doing so further its goals. This effect could lead to public policy that restricts participation in the name of further securing the overall system: a political cycle that could snowball. But what about changing a result? That’s much, much trickier. You can theoretically alter what would have been the actual results by targeting different groups for voter suppression the way robocalls do now. In January 2024, two days before the New Hampshire presidential primary, thousands of likely Democratic voters received robocalls using an AI-generated clone of Joe Biden’s voice. The fake Biden told them essentially to stay home. An independent agent could do more of this, finding its approaches or targets for phone calls, emails or other media itself. An autonomous AI agent might learn that you don’t beat the audit; you attack what the audit assumes are trustworthy. Perhaps it would try to mess with which ballots exist, prevent legitimate votes from becoming paper ballots in the first place, corrupt the records telling auditors where ballots are or interfere with the audit process itself. The underlying idea would be to exploit what the audit is not designed to detect.  A flaw in networked systems, something AI is proving to be very skilled at finding, might be weaponized. Consider Georgia’s system, where the ordering of electronic ballot records potentially exposed how individual voters cast their ballots. One researcher reports using current AI tools to reconstruct the order of 98.9% of the in-person ballots cast in a Georgia election.  Turning these things into an outcome-changing attack would be very difficult even in the AI apocalypse everyone is imagining. If we’re doing what we should, results should continue to be recoverable in a worst-case scenario, even though it might take a little while. That doesn’t mean we should relax, just that we should avoid any dumb mistakes and keep a constant eye on making systems better. Relaxing will never be an option. Read the story on VTDigger here: Where do elections fit in the AI apocalypse?. ...read more read less