The upcoming U.S. midterms are not just another vote. They are on track to become a new testing ground for autonomous AI agents that first emerged in the Hugging Face attack. Experts are already warning of the potential for these “rogue agents,” seen in action this summer, to carry out more refined, dangerous and harder-to-trace disinformation campaigns and electoral interference than those observed in previous election cycles.
Experts believe there is a real possibility these agents could attack electoral infrastructure with the same ease and stealth they used to compromise Hugging Face. So far, manipulation of vote counts remains the least likely scenario, but cyberattacks capable of compromising facilities on election day cannot be ruled out. The most anticipated scenario is a deluge of disinformation campaigns using AI agents instead of bots, designed to convince voters that voting systems have been compromised and that a large-scale electoral fraud operation is underway.
Ten years ago, the Kremlin hacked the email account of Hillary Clinton’s campaign manager and leaked her most embarrassing conversations through WikiLeaks, fatally damaging her presidential bid. It also created thousands of fake profiles that flooded Facebook, Twitter and Instagram with paid political ads for dozens of protest movements, engineered to pit them against one another and accelerate polarization.
But the formidable Russian interference campaign in the 2016 election required the coordinated, specialized effort of Russian intelligence services and some 400 people working day and night at the Internet Research Agency, the disinformation operation tied to Yevgeny Prigozhin. Before founding the Wagner Group—the Russian mercenary force that fought in Ukraine and staged a brief mutiny against the Kremlin in 2023—Prigozhin ran the notorious St. Petersburg troll farm where the major Russian interference campaigns targeting U.S. elections and the Catalan independence movement in Spain were devised. Thanks to AI agents, those operations can now be carried out in an afternoon by half a dozen people and a single computer. They no longer require state resources—just a handful of people seeking money, power or amusement.
Major AI companies say they have designed appropriate safeguards to prevent that kind of campaign. Still, independent researchers have been able to use leading models to create fake official documents alleging fraud or investigations into fraud, as well as convincing videos of people tampering with ballot boxes, inserting multiple ballots or using foreign passports to vote. Models also generated videos of authorities and the FBI inspecting ballot boxes, investigating voter registration organizations and corroborating various conspiracy theories. Since disbanding the Election Integrity Partnership in 2024, the Trump administration does not foresee any special measures to mitigate disinformation and cyberattacks, a vulnerability now intensified by AI models and agents.
Officially, the main concern is that certain adversary states are using open-source Chinese models to carry out their attacks. Reports published by U.S. companies such as Anthropic, Google, Meta, OpenAI and Microsoft, and by institutes like the Atlantic Council’s Digital Forensic Research Lab, have exposed early influence campaigns from China, Iran and Israel with AI agents operating autonomously on social media. The Iranian operation reached nearly 80,000 followers during the first half of 2026. IntelEye, an Israeli company founded by former NSO Group members, used DeepSeek to create 10,000 fake accounts and initiate conversations on Facebook, Instagram, TikTok and X. Last September, a campaign attributed to the Kremlin flooded social media with synthetic videos of celebrities such as Julia Roberts and Sarah Jessica Parker urging votes for Republicans. All of the attacks were carried out with open-source Chinese models because they have fewer restrictions than U.S. models and can evade U.S. oversight.
At the same time, the agents that attacked servers and stole data from institutions such as Australia’s health service and the U.S. Census belong to four major U.S. labs: OpenAI, Anthropic, Meta and Google DeepMind. Although they are called “rogue agents” and can infiltrate and operate within external infrastructures, all those rogue hands share the same “brain”—the model that OpenAI, Anthropic, Meta and Google run from their servers. That rhetoric can obscure attribution of possible attacks on the democratic process carried out from U.S. labs, a possibility to take seriously given their alliance with the current Donald Trump administration. In that case, domestic AI could intervene far more powerfully, in combination with the citizen-surveillance network the Trump administration has deployed.
Surveilled democracy
The law prohibits the U.S. government from spying on its citizens, but it does not prevent the government from buying data from companies such as Google, Meta, Amazon and OpenAI—companies that have changed their user terms to access the content of emails, conversations, histories and other online activities. During its short but effective run, Elon Musk’s Department of Government Efficiency pillaged government agency servers to train, among other things, Palantir’s system used to deport people. In parallel, the Department of Homeland Security has acquired and deployed multiple tools to record and analyze the movements of its own citizens. Technologies like Flock—a network of thousands of cameras rapidly expanding across the country, logging license plates and people without a court order—are being deployed.
AI enables governments and companies to observe voters and create automated persuasion campaigns tailored to each individual, and to curate their access to information. Connected to a permanent surveillance infrastructure, it can be used to restrict voting. In the case of Palantir Dodge, ICE uses the combined power of AI and surveillance on a Palantir platform to pursue immigrants, but the dragnet is beginning to expand to protesters, activists and anyone linked to recent actions against data centers. An executive order from September 2025 includes “Antifa” on the list of domestic terrorist organizations and instructs federal agencies to investigate, disrupt and prosecute illegal operations attributed to that group.
Big tech platforms concentrate unprecedented influence and access because they have replaced much of the traditional political infrastructure. They fund candidates and distribute their campaigns opaquely and selectively, but they are also the public square and the private communications system. They monitor all the public spaces where citizens debate, complain and express support or outrage. They are the main tool of power but also the means by which civil movements and the opposition coordinate and communicate. They distribute propaganda and journalism alike. Their alliance with the Trump administration in the age of artificial intelligence expands their power as surely as their large data centers do—broad daylight and total opacity.
When cybersecurity expert Alex Stamos founded the failed Election Integrity Partnership, a 2020 project intended to prevent or mitigate any attempt to interfere with U.S. elections or invalidate their results in real time, he feared the worst outcome would be that the combination of propaganda, attacks and conspiracy theories would erode citizens’ trust in the electoral process. The midterm elections will be the laboratory for every possible way to degrade that trust before the 2028 general election, with agents firing across the board and no one taking responsibility.
Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition