Anthropic says it disrupted AI-assisted influence network targeting Malaysian voters

Anthropic says it disrupted a commercial, AI-assisted influence operation aimed primarily at Malaysian voters, built around roughly 1,000 fake X accounts, a fabricated news site and profiles of the country's 222 parliamentary constituencies. The case, detailed in the company's September threat assessment and examined in a Guardian report published on 1 October, shows how operators assembled election-focused infrastructure. Anthropic found no evidence that its material broke into authentic online communities, leaving its effect on voters unproven.

What Anthropic found in the Malaysia operation

Anthropic says it removed an account used to run a paid election-manipulation platform and found links to BBS Bilisim Teknolojileri, an Istanbul-based technology company. According to the company, the platform combined automated content production with fake social media identities and fabricated political dossiers. Its report covers misuse it identified and disrupted between December 2025 and August 2026. Those findings describe the operation as Anthropic observed it through its own systems; the report does not establish who ultimately paid for the campaign.

The operators used census and electoral data to prepare profiles for all 222 Malaysian parliamentary constituencies, Anthropic says. Their targeting sought to exploit divisions involving race, religion and royalty. The fake X accounts were configured to inflate engagement and evade detection, while one request sought a million artificial views for the account of Malaysia's sitting prime minister. A request for artificial views is evidence of the operators' aim, not evidence that those views were delivered or seen by voters.

The fabricated outlet, Malaysia Pulse, took reporting from legitimate Malaysian publications, had Claude rewrite it and republished it under invented bylines, according to Anthropic. The company says some material from Russian and Chinese state-aligned outlets was also republished without its original attribution. That process could make recycled or politically slanted material appear to come from an independent local newsroom. Anthropic also says the operators generated dossiers containing false allegations about an opposition politician and civil-society organisations; it found no corroboration for allegations against named people.

What is known about the network's reach

The company says internal detections and recovered indicators helped it map and disrupt the activity. Claude refused or partly refused some requests, including after identifying a fabricated dossier as political defamation, according to Anthropic. The operators then sought less explicit wording while continuing to develop the platform. Anthropic also says they pursued a contract with Malaysia's national communications regulator, but it found no evidence that effort succeeded. These details show both the operators' ambitions and limits to what the report can verify.

Why regional election experts are concerned

In the Guardian's reporting, Katie Harbath, a former Meta elections policy staffer who now leads consultancy Anchor Change, said AI could make political content more personalised, faster and cheaper to produce. She stressed that a campaign still needs a way to distribute it, typically through social platforms. That distinction matters in this case: Anthropic describes a system for making and circulating messages, but does not document a broad real audience for them.

Rossine Fallorina of Philippine thinktank Sigla compared AI-assisted narrative production with a troll factory because it can speed the creation of posts and stories. Bridget Welsh of the University of Nottingham Asia Research Institute Malaysia told the Guardian that such activity feeds on existing political polarisation and divisions. Their comments describe a potential route for harm; neither establishes that the Malaysian platform persuaded voters. Monash University Malaysia lecturer Ben Loh separately described AI-generated material aimed at Rohingya refugees as reinforcing some viewers' existing beliefs, a concern distinct from any measured election effect.

Elections ahead and questions still open

The Guardian reports that Malaysia's next general election could be called for late 2027, while elections are planned in the Philippines and Cambodia in 2028 and Indonesia in 2029. Malaysia's timing remains uncertain. Those forthcoming votes give the techniques described by Anthropic wider relevance, but the report does not show that the same operators will target any future election or that another campaign would gain an audience.

The immediate finding is narrower and concrete: Anthropic says it found and removed a commercially offered system that used its model to help build deceptive political content and an apparent distribution network aimed mainly at Malaysia. Whether any real voters encountered substantial amounts of that content, who commissioned the work and what effect it had remain unresolved in the published account. For readers assessing claims about AI and elections, those gaps separate documented preparation and attempted amplification from a demonstrated change in public opinion or results.