Before the First Click

Investigating illicit networks and malign actors always carries some risk – and increasingly can make researchers into targets themselves. A Bulgarian spy ring targeted former Bellingcat investigator Christo Grozev as retaliation for his Russia research. A private security firm tried to suppress the work of reporters investigating the sexual assault allegations against Harvey Weinstein. Chinese hackers used malware to infiltrate numerous Western news outlets as their journalists investigated Chinese officials. Keeping this in mind, we at C4ADS consider the risks to our analysts and technology before embarking on any investigation. 

Our February 2026 report Deceptive by Design revealed that criminal networks running scam compounds across Southeast Asia are using AI-powered tools to scale their operations and drive cybercrime to new levels of sophistication. When scoping the research that eventually led to the report, we set out to answer three questions: Which AI tools do the scammers in these compounds use? How do these tools build AI models into their products? And who is behind them? 

Answering those questions carried real risk: downloading the scam tools, taking apart their code to see which AI models they used, and searching open sources for information about the companies behind them. But before beginning, we assessed the likely technical risks of pursuing this investigation. When we threat model ahead of an investigation, we are aware that we cannot eliminate all risk, but our goal is to responsibly mitigate as much risk as reasonable. 

Our biggest risk came from the people and organizations we were investigating. Chinese organized criminal syndicates overwhelmingly run scam compounds, and we expected the syndicates to have purchased the tools we studied from opaque companies. The syndicates and the companies controlled the websites and the software we needed to examine, so they could potentially log our visits, identify our devices and network, and deliver malicious code to our machines from thousands of miles away. With that in mind, we identified four main risks: device fingerprinting, network exposure, hostile content, and cross-contamination. 

Device Fingerprinting

A device carries a distinctive fingerprint that includes its operating system, fonts, extensions, screen geometry, and time zone and language settings. Any website that an individual visits can see this information. (Want to try it out yourself? amiunique.org will show you what you’re sharing with the websites you visit.) As a result, every visit to scam infrastructure from that machine could tie an investigation to one identifiable person. 

Organizations trying to evade law enforcement and regulation have exploited device fingerprinting before. For instance, in 2017, the New York Times revealed that Uber used a program called Greyball to show regulators a fake version of its app. The program flagged suspected regulators by their specific devices, payment methods, and other indicators – and then served them the fake version of the app. The same technique could work on researchers.

Network Exposure

An office internet network usually has a fixed Internet Protocol (IP) address, which gets logged whenever someone using that network visits a website. While this can be useful for several reasons, such as seeing where visitors to your website are coming from, it can also pose a vulnerability for individuals conducting research. Traffic from the C4ADS office network says “a nonprofit that publishes investigations” to anyone analyzing the logs. That could potentially expose us in two ways: It could tip off our target about our investigation, and it could also let a hostile actor use that information to scan and then launch an intrusion attempt against our network. 

In one example of network exposure tipping off a target, The New York Times was investigating corruption by former New York State Senate Majority Leader Dean Skelos. In the ensuing trial testimony, a company executive remarked that The New York Times had been on its website “like 12 times that day.” These web logs gave the target advance warning of the investigation.

Hostile Content

As part of our research, we needed to analyze the source code and the live websites associated with these scam centers and had to assume both to be malicious. Whatever they might deploy would land on a machine with the researcher’s accounts, files, and access to our shared drives – all one click away. 

Attackers could exploit this through numerous pathways. One such method is a “watering hole attack.” A hostile actor hides malicious code inside a legitimate website. The site appears safe to most visitors, but serves an infected version to visitors that fit a certain profile, which, in turn, compromises their machines. This means that simply browsing a seemingly trusted page can compromise a visitor’s computer. In 2021, researchers revealed that attackers had injected malicious code linked to Candiru spyware into the Middle East Eye news website, targeting specific visitors to the site. The same campaign also targeted other organizations.

Cross-Contamination

The fourth potential risk was cross-contamination, which occurs when a researcher’s ordinary web browsing and the investigation leak into each other. Most of us have likely experienced this when googling something for a friend and later getting unwanted ads, or seeing someone awkward appear in our LinkedIn visitor analytics. In an investigation, it also runs in two directions. Outbound, a personal Google or LinkedIn session, an autofilled name, or a saved payment card could be visible to the target. Inbound, scam-site cookies, trackers, and ad profiles could follow the researcher home and into personal or work accounts. 

These risks can come up in any online investigation. What changes is how likely a malign actor is to exploit them and how much harm that would cause. 

Our solution always depends on a vulnerability assessment, the specific targets of the investigation, and the potential impact of any specific risk on an individual or our organization. C4ADS uses a range of tools and methods, including virtual machines, air-gapped laptops, and commercial data analysis technology, to mitigate the risks of our investigative work. For this investigation, we determined that organized crime syndicates running the scam centers and their software suppliers posed an elevated risk, and matched our safeguards to that assessment.

In this case, we used a dedicated laptop that only ever connected to the internet via a phone hotspot and was used only for this investigation. Because it was new hardware with no history, it contained nothing that could trace back to our researcher or organization and held no other work or personal accounts. The mobile phone hotspot connection also meant that our office IP address never appeared in any logs. 

We did accept some residual risk: the laptop could become infected; the research target’s logs would still register an increase in visits from a U.S.-based source; a determined adversary could correlate visit timing with our publications; and our process depended on one researcher following these rules without exception. We judged that reasonable for the potential threat.

Our report, Deceptive by Design, examined two sets of AI-powered tools used in Southeast Asian scam compounds: the 007TG suite and KT Smart Translation. Analyzing the tools’ source code, corporate networks, and marketing operations, C4ADS found that AI tools enable scammers to operate globally and at scale, sophisticated AI chatbots likely help scammers deceive targets, and scammers use AI-powered monitoring tools to control a workforce of trafficking victims.

After C4ADS published the report, OpenAI’s fraud investigations team terminated one of the tools’ access to GPT models in March 2026. The GPT-4-powered role-playing chatbot was a core tool enabling scammers of varying backgrounds to convincingly impersonate financial advisors and other professionals in “pig butchering” schemes, in which scammers spend weeks or months building a victim’s trust before persuading them to make a fake investment. Removing this capability degraded one of the most technically sophisticated elements of the scam toolkit.

We partnered with The Associated Press on this work, which led to a report and a PBS FRONTLINE documentary exposing scam center operations.

Earlier in October, C4ADS also partnered with the UK OSINT Community on the 2026 Global OSINT Conference, at which we spoke about our investigative process and the results of this research. 

C4ADS’ mission is to expose the illicit and malign networks that threaten global peace and security. Responsible use of technology helps us protect both the people those networks harm and the analysts who investigate them.