“This is cognitive warfare”: a term leaves military doctrine and enters the political agenda

Von der Leyen tells MEPs that Europe is facing cognitive warfare

Ursula von der Leyen used her State of the Union address on 16 September to call what Europe is facing cognitive warfare. The term is not a loose one. It came out of a NATO study in 2020 and 2021, and Allied Command Transformation has been developing it into a formal concept with more than 20 member nations, covering activity meant to change how whole populations think. It has mostly stayed inside that world until now.

She was blunter about the physical side. Hybrid threats, she said, are less and less hybrid, and the attempted drone attack on Leipzig airport used military-grade material and was carried out by Russian operatives on European soil, which made it an attack on the Union rather than on one country. Her proposal is a European version of NATO’s Article 4: an emergency protocol any member state could trigger for an incident that falls short of armed aggression, obliging the rest to convene and agree a response.

On the information side she pointed to something that already exists, the European Centre for Democratic Resilience, which opened in February for member states, EU officials and independent experts to pool what they are seeing, and said more resources will go into it. She also put the threat on both sides of the border, saying that whether it comes from ultranationalists, anti-Europeans, Russian interference or disinformation, the aim is the same, and that real public concerns are being weaponised to drive a wedge between Europeans, sometimes from outside and too often from inside.

Russia-linked network turns bought celebrity videos into midterm attack ads

Americans vote on 3 November in the midterm elections, which fill every seat in the House of Representatives and a third of the Senate halfway through a president’s term. Seven weeks out, videos of Hollywood actors attacking the Democratic Party began circulating on X, Bluesky and TikTok, stamped with a CNN logo and hashtags such as #AllDemocratsAreCriminals. The actors said none of it: the clips are deepfakes, video altered so that a real person appears to speak words they never did. Researchers attribute them to Matryoshka, a long-running Kremlin-backed campaign, also called Operation Overload, that spreads doctored videos under the branding of real news outlets and pushes them through networks of automated accounts.

One sequence of Julia Roberts came from a 2023 Gucci advertisement, and several clips were bought from Cameo, a website where fans pay actors to record a short personal video message; cloned voices were then laid over the real footage. Emma Caulfield told AFP she felt violated by the use of hers, Ed Begley Jr. and Eric Roberts said the words were not theirs, and CNN objected to the use of its branding. Pablo Maristany de las Casas of the Institute for Strategic Dialogue, a research organisation in London, said the same Cameo trick had already been used in German, Hungarian and Armenian election campaigns.

The posts named candidates in six states, among them Jon Ossoff, a senator for Georgia, and James Talarico, who is running for the Senate in Texas. After the US broadcaster MS NOW showed the videos were fabricated, fresh clips appeared, again branded as CNN, claiming its reporter had retracted the story. Bluesky suspended seven accounts that had posted the clips, all of them taken over from real users; X, where most of the videos were shared, had not acted against the accounts, although some posts have since come down.

Undercover reporter maps how Russia hires amateurs for single jobs in Germany

German broadcaster RTL spent a year inside the way Russia-linked operations hire people in Germany for single paid jobs. The German term for such a recruit is a disposable agent, an amateur paid a small sum for one task at a time, with no lasting connection to whoever is paying and no idea what the task serves. Reporter Angelique Geray set out to be hired as one. She built an account posing as a pro-Russian influencer called Katya, wearing a sailor shirt, praising Vladimir Putin and walking past the German parliament with a Russian flag. Offers of work soon arrived through Telegram, a messaging app, and were paid in the cryptocurrency Bitcoin. The first tasks were trivial: take a photograph, collect a parcel, buy a SIM card. Later ones were not. She was asked to keep watch on an employee of the Christian Democrats, the party leading Germany’s government, to gather information on arms companies and to look into a human rights organisation. Each recruit sees only their own piece, RTL reports.

The group making the offers, which calls itself NoName057(16), was known until now for a narrower trade: flooding websites with traffic until they stop working, a nuisance attack it has aimed at a Bavarian state ministry, the German railway operator Deutsche Bahn and Milan’s airport. German federal police took more than a hundred of its servers out of service in the summer of 2025. What RTL describes is the same group operating in the physical world. The team met Stefan Thumann, a manager at an arms firm that supplies drones to Ukraine, who is in hiding after at least two suspected recruits followed him. They also pieced together a suspected sabotage attempt at Leipzig airport, where parcels rigged to catch fire were placed aboard cargo planes bound for Ukraine. Romania reported something similar on 8 September, when its domestic intelligence service said it had stopped a Russian operation in which a man was photographing Ukrainian cargo aircraft.

The fake influencer account also drew in Jovica Jović, who has around 35,000 followers on TikTok and belongs to the Sahra Wagenknecht Alliance, a German party founded in 2024 that opposes military support for Ukraine. He began advising Katya on what to post. Asked over dinner whether Russian intelligence stands behind his activity, he said that in theory he does not know and in practice he does. Approached later by the same woman, this time as a reporter, he denied any involvement, and his party called him an ordinary member holding no position. Once the account was exposed, a voicemail reached her threatening to shoot her if she ran. Ukraine’s foreign minister Andrii Sybiha pointed to the investigation as evidence of a wider pattern and called for closer counter-intelligence work between European governments.

AI-generated separatist websites target Alberta’s referendum through chatbots

On 19 October the Canadian province of Alberta holds a referendum on breaking away from Canada. Ahead of it, Canada’s National Observer has found 23 websites promoting independence, all written by AI and all sitting on one small web server. They carry names like True Alberta, Farmgate and Vote Yes Alberta, and each is aimed at a particular group of voters: farmers, young people, veterans, mothers, oil-industry workers. Technology reporter Rory White found that those people were not really the intended readers.

The intended readers are AI chatbots. When someone asks an assistant a question, it answers partly from pages it finds on the web, and these pages were built to be among the ones it finds. Their code invites 11 different AI crawlers to come and read, and most of the articles are dated from before the websites existed, which makes them look like long-standing sources.

What they say is wrong in ways that matter. Two of them give the date of the vote as 26 October, a week after the real one. Others list false dates for voting early and misdescribe the wording on the ballot paper. More than 100 Albertans quoted on the sites do not exist, their faces, names and words all generated, among them retired soldiers and Indigenous supporters of independence. A voter who never visits any of these pages can still be handed their claims by an assistant that has read them.

Who is behind it is unknown. On the same server the outlet found a plan for mass-producing websites of exactly this kind, and, on another page, an internal campaign map belonging to Let Alberta Decide, the largest group advertising for independence, listing people who had asked for lawn signs and ranking districts by where canvassing could be done without prompting opponents to turn out. Both pages went offline after the outlet got in touch. The campaign did not answer questions.

An operation like this may not need many people behind it, or any organisation at all: research published in March by Luca Luceri’s team at the University of Southern California, cited in the investigation, found that even small groups of AI agents can run a coordinated influence campaign without a person directing them. Alberta votes on 19 October.

Anthropic says nine influence operations ran on its AI models

Anthropic, the US company behind the Claude AI assistant, said on 10 September that it had shut down nine operations that were using its models to influence public opinion while hiding who was behind them. What the report documents is not only how much text the model wrote, but how much of the work around the text it did.

Take the clearest case. In Bangui, the capital of the Central African Republic, one Russian-speaking man produced the daily output of a radio station, Lengo Songo, in line with what Russian state media were saying, and asked the model to remove the giveaways of AI writing so it would pass as the work of a local reporter. The same account also wrote the station’s employment contracts, which obliged staff to be loyal to the president and to “Russia and its contingent”, and the sheets used to grade their articles. He was running the newsroom and its personnel department at the same time.

That is what the nine cases have in common. Work that used to need an office with staff now needs a few people, or one. A French advertising firm ran around 70 fake news sites in about 20 languages, signed by journalists who do not exist and backed by hundreds of commenting accounts with invented faces, changing political side according to who was paying. In Bangladesh, one man ran 29 accounts for 16 months producing fake political news for village audiences, a case the Dhaka newspaper New Age also reported.

The last case went after individuals. A network tied to an exiled Iranian opposition group copied a real activist’s Telegram account, had the model read thousands of his old messages until it wrote like him, and then held conversations with his contacts, who thought they were talking to him. Its operators had built themselves a tool for this, which they called Viktor. Rather than instructing the AI each time, they kept the instructions in it: words to avoid, sources to trust, how to run accounts without being caught. People who never met produced nearly identical material, because the tool was working from the same notes. Viktor was their own build and nobody else used it, but the habit it represents shows up across the report.

Most of these operations reached hardly anyone, because the company caught them while they were still being put together. The one that did reach people was the radio station, where a real broadcaster carried the material to listeners every day.