Cyber intrusions and disputed information ahead of Taiwan's local elections

Taiwan's National Security Bureau (NSB) sent a written report to the Legislative Yuan on 20 October, and the Liberty Times and Reuters both reported its figures. From January to September 2026, NSB intelligence teams tracking anomalous accounts (social media accounts the bureau considers not controlled by real people, defined in the Perspective below) collected more than 2.5 million items of disputed information, up about 48% from more than 1.69 million a year earlier. According to the Central News Agency, the Government Service Network (GSN), which links government agencies, faced an average of about 2.58 million intrusions a day as of September, with an upward trend since June.

The NSB assesses that China is expanding its use of AI for information manipulation, with the aim of influencing the 28 November local elections. AFP, in a story carried by Hong Kong Free Press, also quoted the report as saying that China's cyber forces use AI agents to scan government websites for vulnerabilities. As of 24 October, we, a community based in Taiwan, found only links to media coverage on the NSB website and not the report itself, so all the figures come from news reports.

The Liberty Times lists Facebook, online forums, Threads and YouTube as the main platforms. Facebook's share rose from 43.6% in the same period of 2025 to 68.9%, and Threads' from 4.5% to 9%. According to the NSB report as relayed by the Liberty Times, the Chinese Communist Party (CCP) mostly has official bodies or state media set the narrative frame, then spreads it through proxy accounts, paid commenters and public relations firms, using content farms, fake websites and AI-generated audio and video.

The NSB director appeared before the legislature's Foreign Affairs and National Defense Committee on 22 October, and the Central News Agency reported six forms of election interference the bureau has identified. They are illicit funding, hosting supporters on trips to China, offering cheap flights so that Taiwanese people living in China can return to vote, spreading disinformation, underground election betting, and hacking specific targets to steal election information. The United Daily News reported that the bureau is tracking about 100 cases.

Perspective

The 2.58 million figure counts "intrusions" (侵擾) on the Government Service Network. How the NSB counts a single intrusion is not explained in any of the reports. The Taipei Times rendered it as cyberintrusion attempts. AFP wrote that Taiwan faced 2.6 million cyberattacks a day, widening the scope to the whole of Taiwan and turning intrusions into attacks.

A separate NSB report in January counted nine categories of critical infrastructure and gave about 2.63 million a day on average for 2025. The two reports cover different scopes, so their figures cannot be compared as an increase or decrease.

The 2.5 million figure is the number of items collected by tracking anomalous accounts. The NSB's term is "disputed information" (爭議訊息), which AFP translated as disinformation disseminated by the CCP. In its January report the NSB defined anomalous accounts as social media accounts not controlled by real users, for example accounts with no personal posts, no social interaction, or a foreign persona posting in Chinese. Whether the October report uses the same definition, and whether the platform shares are counted by items or by accounts, is not explained in any of the reports.

As of 24 October, the report does not appear in the transcript of the 21 October press conference of China's Taiwan Affairs Office of the State Council (TAO). Reuters reported that the TAO did not immediately respond, and noted that Beijing has previously denied interfering in Taiwan's elections or spreading disinformation.

A TAO spokesperson has answered two other questions related to election interference and cyberattacks. At the August press conference, a reporter asked about letters Taiwanese prosecutors' offices sent to village and borough chiefs about election investigations, which mentioned "mainland interference in elections". The spokesperson replied that the Democratic Progressive Party (DPP) slanders the mainland with claims of election interference every time there is an election, for political gain. In September, asked about the Information, Communications and Electronic Force Command of Taiwan's Ministry of National Defense, the spokesperson said the DPP authorities direct the command to frequently attack mainland network infrastructure.

The Australian Strategic Policy Institute (ASPI) analysed Taiwan's 2024 presidential election and observed at least two China-originated threat actors seeking to interfere. Its assessment was that these operations likely had a minimal impact on the integrity of the election results due to the resilience of Taiwan's civil society. According to Radio Free Asia, Doublethink Lab, a Taiwanese civil society group, concluded in its research on the 2024 election that AI-generated material does not by itself establish who created it, and that attribution depends on account networks and posting behaviour.

Under Taiwan's Civil Servants Election and Recall Act (last amended on 1 July 2026, according to the Laws & Regulations Database), prospective candidates and candidates who find deepfake audio or video of themselves between the election announcement and the day before voting can apply to the police for forensic examination at their own expense. In the act, a deepfake is a false depiction of the person's words or actions made by computer synthesis or other technical means. Once the person has the result and submits a request, platforms must restrict access to or remove the content within two days. Spreading false content through a candidate's deepfake with intent to influence the election, where it is capable of harming the public or others, carries up to seven years' imprisonment.

According to a commentary in Tech Policy Press, as of May 2026 no takedown request had been filed with platforms under this provision. Such requests must also be copied to the election commission running the election, and the commentary's author confirmed the zero count with the Central Election Commission (CEC). The author attributes it to the narrow set of protected people and the narrow time window, a deepfake definition that is hard to apply, and the jurisdictional gap with platforms based overseas.

The Ministry of Digital Affairs (MODA) built an election AI deepfake reporting site under the act to help candidates notify platforms. The United Daily News reported on 10 August that it was expected to go live on 25 August, and MODA's statement stressed that the existing legal procedures and responsibilities remain unchanged. In another report on 20 August, the CEC's director of legal affairs explained that most platform operators are overseas and the site makes it easier for candidates to reach them.

In the 10 August report, a Kuomintang (KMT) legislator argued that the CEC is the authority in charge of elections. The legislator questioned whether MODA was expanding its remit and letting the government decide what information gets taken down, and the Taiwan People's Party caucus convener also worried that content unfavourable to the ruling party could be removed as deepfakes. A DPP legislator said the key is whether the content causes harm to the person depicted.

South Korea, Singapore and Japan regulate election deepfakes differently. Under South Korea's Public Official Election Act, as amended on 28 December 2023, no one may produce, edit, distribute, screen or post AI-made virtual sound, images or video that are hard to distinguish from reality for election campaigning, from 90 days before election day through election day. The amendment took effect one month after promulgation, which we calculate as 29 January 2024, since its supplementary provisions give no date. A violation carries up to seven years' imprisonment or a fine of 10 million to 50 million won.

Outside that 90-day window, such content used for campaigning must be labelled as AI-made under National Election Commission rules, and failing to label, on its own, carries an administrative fine of up to 10 million won. We have not looked at enforcement figures.

Singapore's Elections (Integrity of Online Advertising) (Amendment) Act 2024 took effect on 22 January 2025 and amended the Parliamentary Elections Act and the Presidential Elections Act. It is now an offence to publish online election advertising during the election period that realistically depicts a candidate saying or doing something they did not, where the depiction was generated or manipulated by digital means and the publisher knows or ought reasonably to have known that the content is election advertising. The penalty is a fine of up to S$1,000, up to 12 months' imprisonment, or both, and it is a defence to prove that the publisher did not know and had no reason to believe that the candidate had not said or done it.

According to the government's press release on the bill, the ban also covers non-AI techniques such as Photoshop editing, dubbing and splicing, and runs from the issue of the writ of election to the close of polling. We could not find a statutory definition of election period, so that timing relies on the press release. According to the same press release, the Returning Officer can also direct publishers, social media services and internet access providers to take down such content or disable access for users in Singapore. Under the amendment act, a social media service that fails to comply with a direction faces a fine of up to S$1 million, and candidates can ask the Returning Officer to give a direction.

Japan's current Public Offices Election Act has no deepfake-specific provision, and we found no match for terms such as artificial intelligence or deepfake in the current text. The nearest provision makes publishing false matters about a candidate to stop them being elected punishable by up to four years' imprisonment or a fine of up to 1 million yen.

An amendment promulgated on 17 July 2026 takes effect on 1 March 2027 and applies to elections announced from that date. Under it, anyone posting online campaign material with AI-made or AI-altered images or video must make that clear on the recipient's screen, except for minor alterations or images unlikely to be mistaken for real footage. Internet users also have a duty not to publish false matters about candidates so as to harm the fairness of elections, though in the amended text on e-Gov no penalty clause refers to either duty. Large platforms must also take measures to reduce the harm of information that threatens the fairness of elections.

Read against these laws, our view is that Taiwan's deepfake takedown route starts only when a candidate asks for forensic examination, while South Korea bans the use of AI-made content of this kind for campaigning in the 90 days before an election and Singapore lets an election official order removal. Japan's labelling duty will not apply until elections announced from March 2027.

Readers in Taiwan who see a suspicious election message on Facebook, Threads or a LINE group can submit it through the complaints section of the Taiwan FactCheck Center website, or add MyGoPen on LINE by searching for @mygopen and forward it for checking. MyGoPen offers only one-to-one checking and does not handle personal subjective claims. According to the Taichung District Prosecutors Office, as reported by the United Daily News in May, disinformation and foreign interference in elections can be reported by calling 0800-024-099 and pressing 4, and the reporter's identity is kept confidential by law. The Hsinchu District Prosecutors Office lists the same hotline.

The deepfake forensic and takedown procedure under Taiwan's election act is open only to prospective candidates and candidates, and does not apply when an ordinary member of the public is impersonated.

A few questions can help readers weigh the claims made before the vote. First, whether outsiders can see what the NSB's figures and the TAO's counter-accusations are based on, and whether independent research can verify the attribution. Second, whether security and law enforcement agencies can trace foreign money and manipulation before the vote, and whether investigation results are published afterwards. Third, how deepfake takedown rules, in Taiwan and elsewhere, trade off reducing disinformation against protecting speech.