Leadership1 publisher2 min readPublished
Anthropic says Claude AI was used to build a Malaysian election manipulation platform
Anthropic says Claude was used to build a commercial Malaysian election-manipulation platform that ran about 1,000 X accounts targeting 222 constituencies. Writing the content now takes prompts instead of staff, so AI vendors share the policing with the social platforms that still carry the posts.
The Board Room · Leadership desk

What happened
- The operation's fake outlet, Malaysia Pulse, scraped legitimate Malaysian reporting, had the model rewrite it several times and republished it under fabricated bylines.
- The outlet also republished articles from TV BRICS, Xinhua, Sputnik/RIA and CGTN after stripping their attribution.
- The Guardian reports the operation was ultimately ineffective once Anthropic stepped in.
- Anthropic's September threat assessment says actors used Claude to build fake profiles and news sites aimed at audiences on six continents.
- OpenAI banned accounts last year that used AI to generate comments on Philippine politics, many of them backing President Ferdinand Marcos Jnr.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- capability A seller can now offer a buyer constituency-by-constituency messaging across a whole country without hiring the writers a troll farm once needed.
- exposure AI vendors are now the first company able to see election propaganda as it is written, so a missed operation will be charged to them as well as to the platforms.
- constraint A vendor ban removes an operator from one model only; the accounts that carry the posts stay up unless the social platform acts too.
Headcount is the cost that moved in the Malaysian case. "Basically AI is just mimicking what we would have, I guess, termed a few years ago, a troll factory," said Rossine Fallorina of Sigla, a Philippine thinktank that researches AI [8]. "So you don't need 10 people to create a narrative or a story. You can just prompt it and then have it propagated." [8] The Malaysian platform was a commercial product [1]. Anthropic's report lists "private firms selling influence" among the actors it found, alongside "governments, state-aligned propaganda institutions and state media" [5].
Anthropic could act because the writing ran through its own model. The Guardian's account does not say what the intervention involved, or whether X moved against the roughly 1,000 accounts the operation used [1][3].
The skeptic's case comes from Katie Harbath, who worked on public policy for global elections at Meta and now heads the consultancy Anchor Change [6]. "AI can supercharge it in terms of you can make a lot more content that's more personalised, quicker and cheaper," she said. "But you still need a distribution mechanism, and at the moment, that's still your more social media type platforms." [7] I think she is right about reach, and the Malaysian figures fit her. About 1,000 accounts spread evenly over 222 constituencies is roughly four and a half per seat [1]. With so few accounts per seat, the scale sat in the content: the operation pulled census and electoral data to work on fractures of race, religion and royalty [2]. The model's owner sees that writing, and X sees the accounts.
The trade-off for vendors is coverage. OpenAI's Philippine account bans last year [9] and Anthropic's Malaysian intervention [3] each removed operators from one company's model. In my view a commercial seller has every reason to move its buyers to whichever model has not yet banned it [1]. The posts still travel on the platforms Harbath named [7]. Vendor policing then covers production one model at a time.
The election calendar leaves time to plan. Malaysia's general election could be called for late 2027, the Philippines and Cambodia hold polls in 2028, and Indonesia votes in 2029 [10][2]. "It feeds political polarisation and it feeds on divisions," said Bridget Welsh of the University of Nottingham Asia Research Institute Malaysia [13]. The choice in front of vendors and platforms this quarter is whether to build a channel for passing account-level findings between them. Without one, Malaysia's general election becomes the first test of vendor and platform defences running separately [10].
What to watch
- Any statement from X on whether it removed the roughly 1,000 accounts tied to the Malaysia operation.
- Whether Anthropic's next threat assessment names the firm that sold the platform or the clients who bought it.
- Whether AI vendors and social platforms announce a way to share account-level findings before Malaysia's general election is called.