On 10 September 2026, Anthropic reported that an account had used its Claude model in an election-manipulation operation aimed at Malaysian users. According to The Edge Malaysia's account of the report, the operation ran roughly 1,000 fake X accounts and a fabricated outlet called "Malaysia Pulse". It also used Claude to build a voter-profiling platform that ingested census and electoral data for all 222 parliamentary constituencies, targeting race, religion and royalty. Anthropic terminated the associated accounts. The next day, the Malaysian Communications and Multimedia Commission (MCMC) said it takes the report seriously and is examining the issues. It also said it would seek information and decide on follow-up within its jurisdiction. The New Straits Times reported that MCMC said the activities described "do not form part of MCMC's functions or mandate" and that it had no involvement.
The strongest case for the current toolkit
Malaysia's defenders have a fair point. The Communications and Multimedia Act 1998 gives MCMC a long-standing instrument in Section 233, which covers improper use of network services. The Online Safety Act 2025 (ONSA) adds licensed-platform duties, user-reporting mechanisms and codes, including a Risk Mitigation Code and a Child Protection Code, published on MCMC's ONSA portal. A regulator that can compel platforms and prosecute senders has real leverage, and an election-season operation targeting racial and religious fault lines is the kind of harm that justifies state attention. The case for acting is strong.
Where the toolkit does not reach
The problem is structural. These instruments were designed around individual items of content, such as a post, a message or a video, that can be reported, judged and removed. MCMC's ONSA page lists nine harmful content categories: child sexual abuse material, financial fraud and scams, obscene or lewd content, indecent or profane content, harassment or abusive content, violent or terrorist content, self-harm inducement, hate or hostility incitement, and promotion of illegal substances. Coordinated inauthentic behaviour is not on that list. An influence operation of this kind can be made up of individually unremarkable posts. The harm lies in the network: fake identities, laundered sourcing and demographic targeting. Notably, The Edge reports that Malaysia Pulse rewrote local news under fake bylines and republished material from foreign state-aligned outlets with the state attribution removed. Taking down any single post, or prosecuting any single sender, misses that architecture.
There is also a jurisdictional gap. The operators were reportedly linked to a Turkey-based vendor, the accounts sat on a US platform, and the AI tool belonged to a US company. ONSA applies to licensed service providers, and the deeming mechanism covers large platforms. It does not reach a foreign contractor, and it does not cover the model developer. MCMC's statement that the matter falls outside its mandate is therefore less an evasion than an accurate reading of its tools. That is the real finding. Malaysia detected this operation only because a private company chose to publish.
The wrong fix: broader speech offences
The tempting response is to widen Section 233-style liability to cover "AI-generated falsehoods" or political disinformation generally. This would be a mistake. Section 233 has long drawn criticism for vague terms such as "offensive" and "false", and lawyers' groups called in 2023 for its suspension pending reform. Expanding it would hand a broad tool to whoever governs, at the very moment that political-speech credibility is at stake. The report also noted that the actor's pursuit of a contract with the communications regulator was not found to have succeeded, and that Anthropic found no evidence linking the operation to the prime minister or his office. Sorting those claims out needs an evidence-based, neutral process, not a wider offence.
The evidence of harm is also modest. Engagement figures came from the operator's own tools and are unverified, and Anthropic reported no sign of significant traction. Regulation should be proportionate to demonstrated impact, not to the novelty of the technology.
What proportionate looks like
A behaviour-based approach would be narrower and more durable:
- Network-level disclosure duties. Large licensed platforms should report coordinated inauthentic networks affecting Malaysian users to MCMC and publish periodic transparency summaries. That would not depend on a foreign firm's goodwill.
- Precise definitions. Target inauthentic coordination, meaning fake accounts, concealed state or commercial sponsorship and fabricated bylines. Do not target contested opinion.
- Independent review. An agency other than the one running takedowns should assess politically sensitive cases, with published reasons.
- Vendor-side cooperation. Memoranda with AI developers and platforms on threat-intelligence sharing cost little and respect rights. Anthropic's disclosure is evidence that voluntary reporting can work.
- Use existing law first. Where the conduct amounts to fraud, impersonation or electoral offences, the Election Offences Act and criminal law already apply and should be tested before any new speech provision is drafted.
MCMC has said it will decide on follow-up within its jurisdiction. The honest outcome of that review is a recognition that the jurisdiction was drawn for a different problem. Closing the gap should mean giving regulators visibility into networks and keeping the state out of the business of adjudicating truth. The test for Parliament is whether it can build that visibility without enlarging the powers that critics already say are too broad.