Josh Weiner
By Amit Shuster, VP Product, Vetric.io
Anthropic’s threat reports document some of the most advanced attempts to misuse AI, and the September 2026 edition covers a series of cases in which threat actors tried to use Claude for malicious purposes. Two of those cases are influence operations, which Anthropic defines as covert efforts to manipulate public and political discourse while hiding who is behind them, and both produced political content at considerable scale.
The first was run by a French digital advertising agency. It operated around 70 fabricated news sites, each with a matching X account, backed by a network of more than 250 inauthentic commenting accounts. Its articles carried the bylines of journalists who do not exist, and the network published at least 8,913 of them in about 20 languages, shifting its political position to suit whichever client was paying.
The second targeted Malaysian voters through more than 1,000 fake X accounts, managed on a platform linked to an Istanbul-based technology company that sold access to it as a paid service. Inside that operation, Anthropic observed a request for one million artificial views on the sitting Malaysian prime minister’s account.
Both cases show influence being sold as a commercial service, built from fabricated media properties and personas that create the appearance of genuine public support.
AI removed the cost of faking a crowd
The goal of an operation like this is to make a position look more widely held than it really is. Fabricated outlets and fake reporters are built to look like credible, independent sources, so a voter who reads one of these stories would have no reason to suspect that the reporter does not exist or that the outlet was created to influence them.
The targeting can be precise. In the Malaysian case, the operators used census and electoral data to profile voters across all 222 parliamentary constituencies, aiming at the country’s most sensitive faultlines of race, religion and royalty, and they ran a fabricated news outlet, Malaysia Pulse, that republished local reporting and foreign state media as its own. The effect on electoral integrity is harder to measure than conventional fraud, because it could surface much later as lower public trust and participation.
Building this used to take real effort. A persuasive fake persona needed a credible biography and a profile image that could survive a reverse-image search, and a new account needed time to build a believable history before it could influence anyone. That effort worked as a deterrent in its own right, and generative AI has stripped most of it away. A single operator can now produce large numbers of convincing personas, so influence operations can scale with a fraction of the resources they once needed.
Networks give themselves away before they find an audience
Building a network still takes time, and that build phase gives defenders a window to spot coordinated activity before it reaches a meaningful audience. Anthropic reports that most of the content it discovered drew little or no authentic engagement, and in several cases it disrupted an operation before it could build an audience at all.
Spotting a network early depends on seeing how apparently separate accounts and outlets connect, and the cases in Anthropic's report show four signals that give a network away:
Connection and follower graphs. A cluster of accounts created around the same time that mostly follow one another could indicate a coordinated network. This is the signal that turns a list of profiles into a map of an operation, and in the Malaysian case, twelve of the sockpuppet accounts shared a single creation timestamp.
Posting rhythm over time. An account that posts on an unusually regular schedule deserves a closer look, and timing across accounts is just as revealing. For example, the French network's sites published near-identical articles on the DRC-Rwanda conflict within three minutes of each other.
Repeated images and phrasing. Reused or AI-generated profile photos and near-identical wording across accounts and platforms point to a common operator. Many of the French network's commenting accounts used AI-generated profile photos.
Where a claim first appeared. Tracing a narrative back to its origin shows whether it grew organically or was planted. Malaysia Pulse, for example, republished content from Chinese and Russian state outlets with the attribution removed, so it read as independent Malaysian reporting.
Together, this evidence lets an organization assess an influence operation while its reach is still small.
photo credit: istock: IR_Stone
