Anthropic's September 2026 threat intelligence report disclosed a political influence operation in Malaysia that used Claude to build a turnkey election manipulation platform. The system operated roughly 1,000 fake X accounts, a fabricated news outlet called "Malaysia Pulse," and forged documents — all targeting voters across all 222 Malaysian parliamentary constituencies by exploiting racial, religious, and monarchical fault lines. The operation scraped legitimate Malaysian journalism, had the model rewrite it multiple times, then republished under fake bylines. It also laundered content from Russian and Chinese state-aligned outlets — TV BRICS, Xinhua, Sputnik/RIA, CGTN — stripping attribution before redistribution. The core structural shift is labour arbitrage. What once required an army of cyber troops, fake news websites, and coordinated human operators now requires a single person with API access and a prompt. Katie Harbath, formerly of Meta's global elections public policy team, frames it plainly: AI makes disinformation "more personalised, quicker and cheaper." Rossine Fallorina of the Philippine thinktank Sigla puts it more bluntly — AI mimics a troll factory, replacing the ten-person narrative team with a single prompt. The production bottleneck has been eliminated; only distribution remains a constraint, and that constraint is eroding. South-east Asia is the test bed. Maria Ressa has long called the Philippines a "petri dish" for weaponised disinformation, with techniques perfected there later exported to the West. OpenAI banned accounts last year that were generating fictitious social media comments in support of President Ferdinand Marcos Jnr. In Indonesia's 2024 election, an AI-generated cartoon of now-president Prabowo Subianto went viral and materially softened his strongman image. AI-generated content has also been used to incite hatred toward ethnic Rohingya refugees in Malaysia. The threat landscape maps directly onto an upcoming election cycle. Malaysia's general election could be called for late 2027. The Philippines and Cambodia hold polls in 2028. Indonesia follows in 2029. These are nations with predominantly young, hyper-connected populations where social media is the primary information layer and institutional media is either weak, government-aligned, or both. Bridget Welsh of the University of Nottingham Asia Research Institute Malaysia warns that AI "feeds political polarisation and feeds on divisions" — particularly dangerous in Malaysia's ethnically fragmented political landscape. Anthopic says it disrupted the operation and strengthened safeguards. But the structural problem remains unsolved. The company's own report identifies actors including "governments, state-aligned propaganda institutions and state media, as well as private firms selling influence." This is not a rogue actor problem — it is a market. Private firms are now selling AI-powered election manipulation as a commercial service. Anthropic caught one; the question is how many it did not. Harbath raises a second-order threat: flooding the information environment to corrupt the training data of the large language models themselves. In media-thin environments dominated by government-sponsored outlets, a relatively small investment could reshape the information landscape that LLMs draw from. Dr Ben Loh of Monash University Malaysia identifies the psychological mechanism — AI-generated content does not need to persuade; it merely reinforces existing prejudices. "Even though it's AI-generated, to them, it's merely reinforcing what they believe is already true." The Anthropic case was ultimately ineffective because the company intervened. That is the thinnest possible defence. It depends on a single company's willingness and capacity to police its own product — a model that scales poorly, creates perverse incentives, and offers no protection against open-source or self-hosted models that have no corporate moderator at all.
