In mid-August, in the run-up to Latvia’s elections, a lie was carefully crafted about a real military exercise taking place in the southern part of the country.
This was no ordinary lie, destined to bounce around social media accounts in the Baltics. This was a shapeshifter. A disinformation tactic that’s becoming increasingly difficult to defend against, according to Refute, an organisation that tracks influence campaigns.
Refute’s new report (see below) offers a rare anatomy of a shapeshifting disinformation campaign in Latvia, which included a video with at least eight different versions in 15 languages, ranging from Latvian to Russian to Vietnamese. The video falsely claimed that a Ukrainian soldier had shot and killed three NATO servicemen before taking his own life. Refute says all available evidence points towards Russia-linked networks that wanted to deepen divisions and influence opinions about Ukraine and security in the region.
The campaign began its life as a fake breaking news video, styled to look like a genuine report from a respected news channel. It was reworked by local influencers and repeated in Belarus-based broadcasts.
“Three weeks after the false claim was published, the alleged incident was discussed as if it were a fact,” Refute’s report states.
To explore the role these new shapeshifting campaigns are playing in the so-called grey zone of hybrid warfare, Athena spoke to Refute’s CEO and co-founder Tom Garnett.
Elusive targets
Garnett says AI has enabled disinformation campaigns to scale and evolve rapidly. Translations and tailored versions make videos more realistic, he says, while AI agents can be tasked to manage social media accounts and amplify content.
“They [AI agents] are actually being given very high-level objectives,” Garnett explains. “And then it’s a series of accounts working together in a coordinated fashion and managing their own interactions with the population.”
According to Refute’s report, dozens of accounts, with a combined 1.85 million followers, spread the lie about the fabricated shooting during the Baltic Trust 26 military exercise in Latvia.
“That makes it a lot harder to detect, it’s more challenging, the volumes are larger, and therefore from a detection perspective the technology also has to change,” says Garnett.
The objective isn’t just in the Baltics, he adds. It’s about “fracturing societies” around the world. The fake NATO shooting video, for instance, was translated into Italian, Arabic and Spanish.
Garnett says Refute has tracked disinformation campaigns of all kinds, from fake Jeffrey Epstein links to drones damaging fighter jets. Modern campaigns tend to mix traditional propaganda techniques with the latest AI tools. Tracking campaigns in countries like Romania, Estonia and Hungary, Refute is finding fake news websites being created, shapeshifting video content being created with AI tools, amplification through bot farms, and the continued role of human influencers in pushing false narratives into the mainstream.
Reach doesn’t mean success
Garnett says he’s encouraged by some of the resilience Baltic states have shown against sophisticated disinformation campaigns. Latvia’s armed forces rebutted the fabricated NATO shooting within a day. In Estonia, volunteers have helped expose a cloned news website. While this doesn’t necessarily stop a campaign, it can disrupt the link between reach and success.
“Did they [attackers] hit the ultimate overall objectives? Potentially not,” says Garnett. “And that’s largely down to the resilience we’ve seen within those Baltic countries, which are very used to these influence operations.”
Without strong defences, disinformation campaigns can do severe damage. Garnett points to Romania, where the 2024 presidential election was annulled amid allegations of foreign interference and manipulation, forcing a rerun in 2025. Garnett warns that this wider threat extends to companies and individuals.
“I don’t think there’s a country that’s not feeling the effects of hybrid warfare,” he says.
An “arms war to keep up”
In Garnett’s view, when it comes to identifying and disrupting disinformation campaigns, businesses are most vulnerable.
“From boards down to the comms team, they are probably the least prepared.”
While government policy and education (from schools to universities to workplaces) play a big part in defending nations, technology must keep up, says Garnett. He sees opportunity for investors to back innovation in tools that help disrupt campaigns in real-time. This includes developing technology to identify deepfakes, check sources, and trace how content has been altered as it moves between platforms.
For defenders, the challenge is to recognise a campaign beneath its changing forms, and act before its next version finds a new audience.
New report: Democratic Resilience in the Baltics
Cloned news sites, fabricated videos and lies built around real events: Refute’s new report traces Russia-linked disinformation campaigns in the Baltics and how defenders responded. It takes a close look at elections in Estonia and Latvia.
Read more and request a copy of the report here.
Protecting AI’s knowledge well from disinformation poison
Understanding how to defend Large Language Models (LLMs) from being poisoned with false or distorted facts is going to take more data and closer industry engagement.
That’s according to Graham Brookie, senior director of the Digital Forensic Research Lab (DFRLab) in Washington, DC and Atlantic Council’s vice president for technology programs and strategy. Brookie spoke to Athena after attending the recent UN General Assembly, where UK Prime Minister, Andy Burnham, announced plans for a new national centre to fight information attacks.
Brookie says the new “cognitive threat” of LLM grooming is part of a fundamental shift in the information environment and the global security landscape, driven by rapid advancements in AI.
“In 2018, we were talking about bot farms on social media platforms,” he said. “And frankly, that feels a little quaint at this point. Figuring out how to create a research agenda and a policy agenda to solve for some of that is a real challenge.”
According to Brookie, two thirds of Google searches no longer end with a person clicking a link. That means the majority of users are now reading AI summaries and moving on. The question is: where is the LLM getting its information?
Threat actors are changing the “bounds of knowledge” instead of trying to sneak in poisoned ideas, Brookie explains. This contamination of the “universe of sources” the LLMs are trained on presents a major challenge for disinformation defenders. As it gets cheaper and easier to create content at scale, he adds, DFRLab is seeing more experimentation in gaming LLM training.
“We need a lot more data on how LLMs are being poisoned,” Brookie says. “[And] we need more engagement on what best industry practices are, especially for the models that are being manipulated.”
Better sources, stronger defences
Brookie says AI startups fighting disinformation are working to create better, more trustworthy sourcing. Others are giving defenders more powerful tools.
“You can imagine a DFRLab researcher using a team of agents to make their research a lot quicker so that we can expose threat actors more efficiently.”
He’s also noticed a recent increase in sophisticated threat intelligence reporting from frontier labs. This is a major development that positions AI developers on the frontlines of the disinformation war.
Brookie says that, in the absence of regular government reporting, disclosures and reporting from frontier labs support organisations like DFRLab as they adapt their methodologies and set research agendas.
Brookie says leading AI labs are talking about pushing the brake pedal to develop models safely. We need to take advantage of this “increased alignment”, he says, to boost information resilience.
News: what changed this week
Influence campaigns are manufacturing credibility and blending digital deception with real-world operations. OpenAI says it banned accounts linked to elaborate Russian and Iranian “false front” operations that used ChatGPT models to “launder geopolitical, conflict-related messaging”. The Iranian operation used “a stable” of seven fake journalist personas to place almost 100 articles across online publications. The Russian operation recruited unwitting researchers to staff a real world think tank in Latin America.
Disinformation discovered in six out of 10 elections, a new global report finds. International IDEA’s review of 2025, produced with the Electoral Integrity Project, recorded disinformation and misinformation in at least 62% of the 39 national elections it monitored. The danger extends beyond misleading voters about candidates.
Read International IDEA’s report
But…
Reach alone won’t tell us how much harm disinformation can do. Full Fact and the University of Westminster assessed dozens of examples of AI-generated or altered misinformation - and found that less than half had substantive potential for specific harms (including financial loss, damage to health or social unrest). The report says we need to better understand the difference between “the potentially harmful and the inconsequential”.
Europe puts its disinformation defences to the test. A new EU-funded project, FIMI-RESIST, will assess 20 existing tools and test their usefulness with journalists, fact-checkers and other defenders. Led by the University of Mainz, with partners including Factiverse, the €3.5 million project addresses a practical gap: establishing which tools work, for whom, and against which threats. (For more on Factiverse and its work see the key takeaways from Athena’s panel in Copenhagen.)
Read the University of Mainz announcement
Solutions: make life difficult for threat actors
Tom Garnett is watching two companies - Reality Defender and Ghost Shift - working on different parts of disinformation defence: spotting fabricated content and disrupting those behind attacks.
Reality Defender detects AI manipulation in audio, video and images, including during live calls and meetings. Its tools help organisations flag cloned voices and fabricated footage before they trigger a payment or enter a news report. That can help stop impersonation scams and expose fabricated material used to damage reputations.
Ghost Shift is a consultancy firm specialising in gathering online intelligence and helping clients disrupt harmful campaigns launched by adversaries. It also specialises in counter-narratives. For those tackling influence operations, it addresses a key challenge: moving from watching a campaign spread to making it more difficult for the threat actor to run it.
Full Fact offers another part of the defence: checking claims and helping reliable information reach the public. It’s also developing a benchmark to assess the accuracy, transparency and consistency of Large Language Models. Its latest report, covered in our news section above, examines which AI-generated falsehoods have the potential to cause harm.