An election manipulation platform targeting Malaysia: GTG-84005 case

What happened

We identified and removed an account that used Claude to operate a commercial election manipulation platform primarily targeting Malaysian users. The network consisted of about 1,000 fake X/Twitter social media accounts, a fake news outlet, and a series of fabricated dossiers.

The platform was disguised as a defensive cyber intelligence and counter-disinformation tool. However, our investigation found a clear link to BBS Bilisim Teknolojileri, an Istanbul-based technology company that sold access to the platform as a paid 'influence-as-a-service' capability. According to the threat actor's own documentation, the infrastructure was marketed as a 'military-grade, AI-driven, real-time political action ecosystem.'

The operation involved the actor using the platform built through Claude to profile and target Malaysian voters constituency by constituency, using the census and election data it ingested. The platform managed a network of about 1,000 fake accounts optimized to inflate engagement metrics and evade social media platforms' detection systems. The setup also ran a fake news website called 'Malaysia Pulse,' fed by an AI rewriting pipeline for this synthetic news outlet.

The people behind the operation also generated fabricated intelligence dossiers to spread false allegations against an opposition politician and civil society organizations. These allegations were entirely fabricated by the actor.

We assessed the operation as Tier 2 on the Breakout Scale, meaning the assets were distributed across multiple platforms but there was no evidence they broke out into real communities.

What the AI did

The actor used Claude Code to build custom dashboards for managing, running, and tracking the fake account network. These dashboards tracked metrics such as the likes and views that deceptive accounts generated for each target. One example was an account of a senior Malaysian government official recording figures in the millions. Since these figures were self-reported by the actor's own tool, we could not independently verify them.

The operation used real census and election data, as well as millions of voter records, to target the country's most sensitive political and social fault lines: race, religion, and royalty across all 222 parliamentary constituencies in Malaysia.

The platform managed over 1,000 fake X/Twitter accounts. Each account had 'account-aging' logic to make it look authentic for a period before being deployed in influence operations, including regularly updating cookies and IP addresses. The dashboard included a parameter allowing users to adjust how many total artificial views each target should receive. We observed a request, in support of Malaysia's incumbent prime minister, for 1 million artificial views for his account.

The actor generated allegations against named individuals that our model's own research could find no corroboration for.

In cases where Claude refused to perform the operations the operation requested — including after it identified a document as political defamation material — the actor negotiated the use of sanitized wording to continue building the same capability.

The actor sought a contract with Malaysia's national communications regulator. We found no evidence this attempt succeeded.

Claude was used to build the constituency targeting system using real election data, to design and run the fake social media account network and its detection-evasion logic, to rewrite and launder fake news, and to iterate on fabricated dossiers.

The synthetic news outlet also scraped legitimate Malaysian reporting and had the model rewrite it multiple times before republishing under fabricated bylines. It republished articles from Russian and Chinese state-aligned foreign media, including TV BRICS, Xinhua, Sputnik/RIA, and CGTN, stripping state attribution to make them look like local independent Malaysian reporting.

What the report observed

The report confirmed the platform's purpose, population classification method, and the existence of stockpiled fake accounts. The report confirmed the operation was discovered and banned during its pilot phase. The report confirmed its observations closely match the capabilities described in Forbidden Stories' 2023 investigation.

The report lists relevant indicators of compromise (domains, IPs, etc.) at the end for defenders to investigate. Per this site's safety boundaries, specific indicator values are not reproduced.

Confirmed & unknown

Platform response

Anthropic says it identified the account through internal detection, mapped the operation's full footprint, and used recovered indicators to block future abuse; the original report shares domain, IP, and account indicators with industry partners.

Limits of response：The action targeted the platform account and known indicators; the report does not state whether the commercial platform still offers similar capabilities to other clients.