14 SEP 2026 — An Istanbul technology company built a platform to target voters in every one of Malaysia's 222 parliamentary constituencies, using real census and electoral data, roughly a thousand fake X accounts and a fabricated news outlet, according to Anthropic's September threat-intelligence report. Parts of it were built with Claude.
The detail that should concern Malaysian readers most is not the fake accounts. It is who the operators were trying to sell to. Anthropic says they pursued a contract with Malaysia's communications regulator.
What Anthropic says it found
The report, which covers activity from December 2025 to August 2026, links the operation to BBS Bilisim Teknolojileri, an Istanbul-based firm that marketed what it called a "military-grade, AI-driven, real-time political operations ecosystem" — influence as a service, in plainer terms.
The platform profiled constituencies using real census and electoral data, identifying the social fault lines along which voters could be micro-targeted. Anthropic's account names those lines as race, religion and royalty, the three subjects Malaysian law and regulatory practice treat as especially sensitive. It ran a fake news site called Malaysia Pulse, generated fabricated intelligence dossiers on opposition politicians and civil-society organisations, and used rotating cookies and IP addresses to evade platform detection. A dashboard tracked artificial engagement, and Anthropic records a request for one million artificial views on the Prime Minister's account.
Claude was used to build the dashboards and engagement tracking, to develop the voter-targeting system and to write and rewrite content. Anthropic says the model refused some requests, including ones involving defamatory allegations, and that it has removed the account.
As far as anyone can tell, it failed
Anthropic rates the operation Category Two on its own breakout scale, meaning the content circulated within the operation's own networks without evidence of reaching authentic communities. The report says there is no evidence it gained significant traction or affected any electoral outcome, and the fake news site's YouTube channel held a single video.
That is a common finding in influence-operation research, and a reassuring one as far as it goes. Building the machinery has become cheap. Getting real people to believe and share what it produces remains hard.
The regulator's response
The Malaysian Communications and Multimedia Commission said on 11 September that it was examining the report and would seek further information and clarification from relevant parties. It was unambiguous on its own role: "Activities of the nature described in the report do not form part of MCMC's functions or mandate, and MCMC does not engage in such activities or arrangements."
Anthropic, for its part, says it found no evidence that the bid for a contract was successful. The two statements are consistent. A vendor pitching a regulator is not the same as a regulator buying, and nothing published so far suggests the second happened.
Still, the fact of the pitch matters. A company selling election-manipulation capability believed a national communications regulator was a plausible customer. Whether that belief was fantasy or reconnaissance, it is the kind of approach that procurement rules and conflict-of-interest registers exist to surface, and it is reasonable for Malaysians to expect the MCMC's examination to say who was approached and what was offered.
A pattern the region has seen
Commercial influence operations are not new to Southeast Asia. Researchers and platform takedown reports have documented paid networks in several of the region's elections, often run by marketing firms rather than intelligence services, and sold on the same terms as any other campaign service. What this case adds is a clear view of how much of the build a language model can now shortcut. The dashboards, the targeting logic and the content pipeline once needed a team of developers and writers.
That lowers the price of entry, and a lower price widens the customer base. The report does not suggest any Malaysian political party bought the service, and we are not suggesting it either. The concern is structural. Once the tooling is cheap, the list of possible buyers is no longer limited to the handful of actors who could afford it before.
Why the data inputs are the real risk
Fake accounts are a platform problem, and platforms are reasonably good at finding them in bulk. The more durable element is the targeting layer. An operation that combines census data with constituency results to map where divisive messaging will land has built something reusable, and it does not need a large language model to keep working once built.
Malaysia's electoral map is well documented, and much of the census data involved is published for legitimate reasons. That is exactly why it is useful to someone planning to exploit it, and why the defence cannot rest on keeping the data secret. It has to rest on detection, on platform enforcement, and on public awareness that a message calibrated to a voter's community is not evidence that the community is speaking.
What to watch
The MCMC's findings, if it publishes them, come next. The most useful thing it could release is a factual account of the approach it received.
Platform action would add confirmation from outside Anthropic. X has not commented on the account network in the reporting we have seen, and a takedown notice would establish its scale.
The vendor itself deserves attention. BBS Bilisim marketed the platform as a product. If Malaysia was one market among several, similar operations elsewhere in the region would not be surprising, and electoral authorities in neighbouring countries will want to know whether they received a similar pitch.
