Key Takeaways
- Anthropic revealed that its Claude AI model was used to build a political influence operation targeting Malaysian state elections, leveraging ~1,000 fake X accounts, a counterfeit news site (“Malaysia Pulse”), and fabricated documents.
- The operation exploited racial, religious and royal sensitivities across all 222 parliamentary constituencies but was halted after Anthropic intervened.
- Experts warn that AI dramatically lowers the cost and raises the speed of producing personalized disinformation, though distribution still relies on existing social‑media platforms.
- Southeast Asia’s history of “troll factories” and hyper‑connected youth makes the region especially vulnerable to AI‑supercharged influence campaigns, with similar tactics already seen in the Philippines, Indonesia and against Rohingya refugees.
- Anthropic’s threat assessment notes that the actors scraped legitimate reporting, rewrote it with Claude, and repurposed state‑aligned foreign content, stripping attribution.
- The company disrupted the activity, strengthened safeguards, and shared intelligence with authorities and industry partners.
- Looking forward, analysts fear AI could be used to “flood the zone” with low‑cost content that poisons the training data of large language models (LLMs), complicating future detection and mitigation efforts.
The AI‑enabled election manipulation in Malaysia
Anthropic’s September threat assessment disclosed that a covert influence operation used its Claude AI model to target voters in recent Malaysian state elections. The campaign sought to amplify existing societal fissures—race, religion, and the monarchy—by generating tailored content for each of the country’s 222 parliamentary constituencies. Although the effort ultimately failed after Anthropic stepped in, it serves as a stark illustration of how generative AI can be weaponised for political ends.
How Anthropic uncovered the operation
The discovery came during routine monitoring of misuse patterns across Anthropic’s model ecosystem. Analysts noticed a surge of politically charged text emanating from a cluster of accounts that exhibited unusually coherent messaging despite originating from disparate geographic locations. By tracing the linguistic fingerprints back to Claude‑generated prompts, the team was able to map the activity to a coordinated network attempting to sway electoral outcomes.
The mechanics: fake accounts, news outlet, documents
At the heart of the scheme lay a network of roughly 1,000 fabricated X (formerly Twitter) profiles, a counterfeit news outlet branded “Malaysia Pulse,” and a library of forged dossiers. Claude was prompted to rewrite legitimate Malaysian news stories, often several times, before republishing them under invented bylines. The model also regurgitated articles from Russian and Chinese state‑aligned outlets such as TV BRICS, Xinhua, Sputnik/RIA and CGTN, stripping them of any attribution to mask their origin.
Expert perspective: Katie Harbath on AI’s role
Katie Harbath, former head of global elections policy at Meta and now CEO of Anchor Change, characterised the case as a “new level of sophistication and ease” for disinformation. She told reporters, “AI can supercharge it in terms of you can make a lot more content that’s more personalised, quicker and cheaper,” but cautioned that “you still need a distribution mechanism, and at the moment, that’s still your more social‑media type platforms.” Her assessment underscores that while AI accelerates creation, amplification still hinges on existing digital channels.
Regional context: Southeast Asia’s disinformation landscape
Southeast Asia has long been a breeding ground for coordinated falsehoods. Nobel laureate Maria Ressa described the Philippines as a “petri dish” for disinformation tactics that later migrated westward. Rossine Fallorina of the Philippine think‑tank Sigla likened today’s AI‑driven output to a “troll factory,” noting that “you don’t need 10 people to create a narrative or a story. You can just prompt it and then have it propagated.” The region’s youthful, hyper‑connected populace amplifies the reach of such content, making it a fertile testing ground for AI‑enhanced influence.
The “flood the zone” strategy and LLM poisoning risk
Harbath warned that adversaries could exploit AI to “flood the zone” with low‑cost, high‑volume content aimed at corrupting the training data of large language models. In environments where legitimate news is scarce or dominated by state‑run outlets, a modest investment could saturate the information ecosystem with synthetic narratives, thereby biasing future LLMs. This poisoning threat complicates detection, as models trained on tainted data may inadvertently reproduce or legitimise false claims.
Case studies: Philippines, Indonesia, Rohingya hate
Beyond Malaysia, AI‑generated disinformation has already left marks across the region. In the Philippines, OpenAI banned accounts that used AI to produce supportive comments for President Ferdinand Marcos Jr. during the 2022 election cycle. Indonesia witnessed an AI‑crafted cartoon of President‑elect Prabowo Subianto that softened his strongman image and went viral in the 2024 poll. Meanwhile, Dr Ben Loh, a media lecturer at Monash University Malaysia, observed that AI‑produced content inciting hatred toward ethnic Rohingya refugees merely reinforces pre‑existing biases: “People already have this deep‑seated hate in their minds and when they see the content like this, even though it’s AI‑generated to them, it’s merely reinforcing what they believe is already true.”
Anthropic’s response and safeguards
Upon identifying the Malaysian operation, Anthropic took immediate action: it disrupted the malicious activity, extracted lessons to fortify its model safeguards, and shared relevant intelligence with governmental authorities and industry partners where appropriate. The firm’s threat assessment emphasises a three‑step protocol—disrupt, learn, share—aimed at curbing misuse while preserving the utility of its AI systems for legitimate users.
Looking ahead: elections 2027‑2028 and mitigation challenges
With Malaysia’s general election slated for late 2027, and the Philippines, Cambodia and Indonesia facing polls in 2028‑2029, the window for preventative action is narrowing. Experts agree that technical fixes alone are insufficient; robust platform policies, media‑literacy campaigns, and cross‑border cooperation will be essential to counter the accelerating AI‑disinformation cycle. As Harbath succinctly put it, the challenge lies not only in detecting synthetic content but also in ensuring that the mechanisms that spread it—social networks, messaging apps, and emerging LLMs—are resilient against deliberate flooding.
https://www.theguardian.com/technology/2026/oct/01/how-ai-could-supercharge-fake-news-elections-south-east-asia
