Executive Summary

Spam texts that people in Korea reported to the government have ended up inside commercial blocking services, and the effect has now been published as a count. The figures came out at the 2026 Spam Data Opening Results Briefing, held in Seoul on November 20 by the Korea Internet & Security Agency (KISA) and the Korea Media and Communications Commission (KMCC). News that public data has been opened is common. News that follows the data into a named service and counts what it stopped there is not.

The largest number is 500 million. That is how many malicious texts Samsung Electronics' AI-based malicious message blocking service has stopped since it launched in March 2025, running on the spam report data KISA hands over. On the voice side, LG Uplus's AI call service ixi-O blocked about 12 million spam calls in the three months since that partnership began. The two figures cover different stretches of time and different media, though, so reading them as one combined result is a stretch.

Sections 1 to 4 set out what the KISA and KMCC announcements and the coverage confirm. Section 5, which reads the case as a question about AI-Ready Data, is this article's own.

Key figures

Sources: KISA press release (November 20, 2026) · KMCC press release (November 16, 2026) · ZDNet Korea · Jaekyung Ilbo

The first two cells count what was stopped inside services the opened data reached. The third is the supply flowing into one of those services, and the fourth is the broader measure of how much spam people actually received over the same period.

About 500 million

Malicious texts stopped before delivery

Samsung's service, cumulative since its March 2025 launch

About 12 million

Spam calls blocked in advance

LG Uplus ixi-O, over three months of the partnership

500,000 a month

Data Samsung receives

Sender numbers, risky links, message text and the like

8.52 messages

Spam received per person a month

First half of 2026, survey of 3,000 people, down 17.7%

1

The Text Never Shows, the Call Never Rings

There are two moments at which spam can be stopped. One is after it arrives, when you delete it. The other is before it arrives, when it never gets through. What KISA presented this time belongs to the second kind. None of the items in the count ever showed up on a screen, and none of the calls made a phone ring.

The AI-based malicious message blocking service Samsung Electronics runs with KISA launched in March 2025. It picks out texts that look like illegal gambling, loan offers, stock tips or smishing, and it judges them against KISA's spam report data. Since launch, the service has filtered about 500 million malicious texts before delivery.

The work with LG Uplus started in July 2026 and covers calls rather than texts. KISA shares voice spam data to sharpen how accurately the AI call service ixi-O identifies spam calls, and the spam sender numbers ixi-O pulls out of call analysis are wired back into KISA's real-time blocking system. In the three months after that work began, about 12 million spam calls were blocked before they rang.

Several outlets added the two into a single figure of 512 million blocked. The arithmetic holds, but one side covers 20 months and the other three, and one counts texts while the other counts calls. They read better apart.

Three items were on the agenda that day: the state of illegal spam in 2026, the progress of KISA's spam data opening, and the cases and results of using that data by sector. The first two are close to annual reporting. The third is the part that came with numbers attached this time.

2

One Report Becomes One Label

KISA's job in this program splits in two. First it collects the spam that the public reports, then it classifies and analyzes that pile automatically with AI. Opening comes after that, when the classified data goes out to public bodies and to companies.

The private recipients named in the press release are mobile carriers, handset makers and others like them, and two of those have blocking counts attached to them in this announcement. The one case where the supply volume is spelled out is Samsung Electronics. Samsung receives sender numbers, risky links and message text from KISA and uses them to improve its blocking, and the volume handed over runs to about 500,000 items a month. Over a year that is a six-million-item sample of current spam flowing into one company's classifier.

What deserves more attention than the volume is how this data was made. A spam report is a record of an actual recipient marking, on their own phone, that this one is spam. It was not produced by commissioning a labeling job. It is a label applied by the person who was hit, at the moment they were hit. Collection and labeling finish in a single motion.

Timing matters too. Spam sender numbers and short URLs get swapped out within days. A list gathered last year will not stop what arrives today, which is why a supply that keeps coming month after month is worth more than one large dataset handed over once. Here the value of opening comes from the interval rather than the volume.

3

One Way Out, and a Way Back

The same opening program connects the two companies differently. On the Samsung side, KISA gives and the company receives, and that is the whole path. On the ixi-O side there is one more leg, a return. The spam sender numbers ixi-O extracts from calls are fed into KISA's real-time blocking system, so a company's analysis refills a public blocklist.

Same reports, two ways to connect Samsung Electronics — one direction Citizen reports marked by the recipient KISA AI sorting about 500,000 a month Samsung message blocking about 500 million blocked to date LG Uplus ixi-O — both ways KISA blocking system voice spam data shared better detection numbers returned ixi-O call analysis 12 million blocked in 3 months
▲ Pebblous original diagram, arranging what KISA announced by direction of flow. The announcement does not break the counts down by contribution. | Source: KISA press release (November 20, 2026)

Once a connection runs in the opposite direction, the supplier gets to see the outcome of its own data. How often the numbers a company sends in are caught across the whole blocking system accumulates at KISA, and that record becomes material for the next round of classification. Because the receiving company's judgment comes back to the body that handed the data over, opening does not end with a single handover.

This give-and-take did not begin with ixi-O. KISA signed an agreement of the same shape with KT in September 2024. KISA passes spam report data to KT, KT's AI Clean Messaging analyzes it and pulls out sender numbers, and KISA takes those numbers back, puts them on a blocklist and pushes the list down to every text relay operator. Bundling heavily and repeatedly reported numbers into a blocklist for relay operators was something KISA had been doing well before that. What joined the list in October 2024 was numbers a company had found through its own analysis.

Still, it is early to credit that structure for 12 million in three months. Voice and text do not move in the same daily volumes, and a call leaves time for analysis while it is connected, whereas a text arrives all at once. The announcement carries no material setting the conditions of the two projects side by side. What is confirmed is that the connections differ and that blocking counts came out of both.

4

The Police and the Financial Watchdog Get the Same Reports

Companies are not the only recipients of the spam data opening. The Korean National Police Agency, the Financial Supervisory Service, which is the country's financial watchdog, the National Gambling Control Commission, the Ministry of Gender Equality and Family and the Ministry of Food and Drug Safety take the same report data and use it in their own areas. The actions split into three kinds: feeding investigations, suspending phone numbers, and blocking malicious URLs. One press of a spam report button, and that record branches at once into a company's classifier and an investigator's lead.

Recipients are many because illegal spam now serves many purposes. KISA says illegal spam is being abused as a vehicle for crimes including illegal gambling, illegal finance, illegal medicines, illegal adult material and phishing. Gambling ad texts are a matter for gambling oversight and medicine ads are a matter for the food and drug ministry, so splitting one classified pile by jurisdiction follows naturally. The branches are not the same size. The statistical annex to the release the KMCC issued on November 16 puts the advertising categories in first-half 2026 text spam at 40.5 percent gambling such as lotteries, 15.4 percent finance in the form of investment pitches, 4.8 percent illegal loans, 3.3 percent adult material and 36.0 percent other. The largest single share goes to the National Gambling Control Commission.

Enforcement tightened over the same period as the opening. In June 2024 the KMCC introduced a transmission eligibility certification scheme requiring text resellers that run bulk internet messaging to be certified by a text relay operator before trading. In September 2026 it approved an amendment to the enforcement decree of the Network Act imposing fines of 1 to 6 percent of related revenue, or 100 million to 2 billion won, on illegal spam senders and on operators that neglect their duty to prevent it. Until then the ceiling was an administrative fine of 30 million won.

The numbers moved in different directions by channel. The first-half 2026 spam distribution report that the KMCC and KISA issued separately on November 16, four days earlier, puts total reports and detections at 21.46 million, down 9.3 percent from 23.67 million. Within that, text fell 45.9 percent from 12.88 million to 6.97 million, while voice rose 28.1 percent from 8.73 million to 11.18 million and email rose 60.7 percent from 2.06 million to 3.31 million. What users feel, on the other hand, fell across all three types: spam received per person a month was 8.52 messages, 17.7 percent lower. The same document attaches a caveat to that mismatch, noting that the volume received is a survey result while reports and detections are KISA's own report and detection counts, so the two trends can differ. The first comes from asking 3,000 people aged 12 to 69 in May. The second is what actually reached KISA or was caught by it.

Change in spam reports and detections by channel First-half 2026 spam distribution report figures (%) Total -9.3% Text -45.9% Voice +28.1% Email +60.7%
▲ Pebblous original diagram. Total reports and detections fell while voice and email rose — the KMCC's annex names tighter text spam regulation as the driver behind the voice surge. | Source: KMCC press release (November 16, 2026)

The side that rose is not treated as a coincidence by the bodies that published it. Next to the jump in voice spam reports and detections, the annex notes that tighter text spam regulation drove the surge in voice spam reports. The reading is that the force squeezing text pushed the traffic toward voice. Voice is exactly where the ixi-O partnership sits, and voice is where the KMCC and KISA have put their next move: a plan to work with mobile carriers so that the spam report function in voice call apps connects to KISA's reporting system. Rather than reworking the blocking model, that choice widens the door reports come in through.

The announcement leaves the decline undivided between the data opening and tougher enforcement. All that is clear is that the two worked over the same period.

Shin Dae-gyu, head of KISA's digital infrastructure division, described the spam data opening as a cooperation model that blocks illegal spam effectively by combining the agency's AI analysis technology, the response capacity of specialist bodies in each field and the technical strength of private companies. He said KISA would keep widening its work with partner agencies and companies and keep advancing its AI-based analysis and its data opening system, so as to produce reductions in illegal spam that the public can feel.

5

Why Pebblous Is Watching This Case

The measures that come up most often in writing about open public data are how many datasets were opened and how many records were posted. Records that follow what changed after the opening are much rarer. That gap is what this blog looked at in the hole in public data evaluation and in collection without a standard. This announcement stands out because it runs the other way. Which service the data went into, and what was caught there and how much of it, came out together.

The time it took to reach these numbers was not short. The Korea Communications Commission, the KMCC's predecessor, said it was pushing ahead with a spam big data opening project in January 2020, and the starting point was handing spam data to the Korea Racing Authority in October 2019 to block illegal horse-racing sites. Even then the plan said the data would go beyond regulators to solution developers and universities for research on spam statistics and blocking technology, and it included automating a sharing procedure that was being run by hand. Widening the set of recipients and turning that channel into a system took the years in between.

Three properties of this data can be picked out. First, the labels cost nothing. The moment a recipient who was hit presses the report button, a verdict of spam is attached to the data, so no separate labeling bill arises. Second, it does not stop. 500,000 a month is not a block received once but a flow arriving every month, and in a domain where sender numbers change within days, that interval matters more than the volume. Third, more than one party uses it. The same classification result serves a company's model and an investigator's lead at the same time.

None of the three is unique to spam. A worker's report of an equipment fault on a factory floor, a damage log in logistics, a correction of a false positive in service operations all have the same shape. A person has already attached a verdict in the course of doing the work, the verdict is current, and there is more than one place to use it. Yet records like these usually stay inside one department's system.

A team that holds data can put these questions to its own records.

  • Where are the records that already carry someone's verdict? Report, rejection, correction and rework logs are candidates. That is the place to count before ordering any labeling work.
  • Does the interval at which those records arrive keep up with the speed of the problem you want to solve? A dump that comes down once a quarter cannot follow something that changes weekly.
  • Does what the recipient learns ever find its way back to the source? As the ixi-O case shows, that connection makes the next round of classification that much better.

If AI-Ready Data means data in a state a model can use directly, this case shows that the state does not come only out of expensive processing. Citizen-reported spam is not a clean benchmark dataset. It still carries labels, it is current and its use is obvious, so the companies that received it could put it into a model straight away. Which of the records still closed in our own field already meet the same conditions is the question this announcement leaves with the reader.

Editor's Note

The figures in this article were checked against KISA's press release, coverage published the same day, and the first-half 2026 spam distribution release the KMCC and KISA put out four days earlier. The supply volume of about 500,000 a month is not in the body of a press release; it appears in an article covering the event. The advertising category shares and the line about tighter text spam regulation driving the surge in voice spam reports are not in the body of that release either; they were read off the statistical charts attached as an annex. The blocking counts are tallies by KISA and the participating companies, with no external verification published and no definition given of what counts as one item, whether a send attempt or an intended recipient. How much of ixi-O's 12 million is owed to KISA's data is not in the announcement either. The reading of industrial data in Section 5 is this article's judgment and is not in KISA's announcement.

R

References

Official announcements

News coverage