Artificial intelligence (AI) has slashed the time required to prepare phishing campaigns from 16 hours to five minutes. In just ten seconds, AI can aggregate information about an individual from over 50 sources and craft a personalized fraudulent message.
By accelerating reconnaissance, AI enables attackers to launch “hyper-personalized” campaigns. The time needed to prepare phishing emails has dropped from 16 hours to five minutes, while the setup time for phishing websites has been reduced to 30 minutes. Analysts at the Solar Group (GK Solar) estimate that AI-driven phishing has a 54% success rate, compared to 12% for manually crafted phishing attempts.
It is worth noting that the number of phishing websites is not a reliable indicator of the total volume of personalized attacks. A single site can be used to target many individuals, and some schemes operate entirely without a website—relying instead on direct correspondence or phone calls. Furthermore, a single domain can host numerous subdomains, each featuring personalized pages tailored to specific targets.
“AI lowers the cost of crafting convincing messages and allows for large-scale production. Generative tools help attackers draft communications more quickly, eliminate linguistic errors, and create content for targeted phishing,” explains Alexander Vurasko, Director of Business Development for Solar AURA (a center for monitoring external digital threats within the Solar Group).
According to Vurasko, the most sought-after data is information that helps establish a credible opening for a conversation—such as a person’s name, phone number, address, place of employment, or social circle. For instance, in “fake courier” scams, fraudsters might cite a victim’s name and delivery address to convince them that they are speaking with a legitimate representative from a store.
The next level involves leveraging context to select the most effective “cover story” or pretext for the scam. Details such as a person’s workplace and the name of their manager can be used to simulate a work-related assignment; information about relatives can facilitate a message appearing to come from a loved one; and posts about personal interests can be leveraged to offer a relevant product or service. For instance, companies have documented attack scenarios based on prior research into a potential victim’s social media, social circle, professional life, and financial situation. Thus, personalization of this kind existed long before the current proliferation of generative AI.
“It would be a mistake to assume that all such information comes exclusively from open sources. Data from social media can be supplemented by information from leaks, the contents of compromised accounts, or details a person may have shared in a previous conversation. Data leaks are a significant resource for preparing attacks,” the expert notes.
Neural networks can aggregate information about an individual from 50 different sources in just ten seconds.
Fraudsters often do not need a detailed psychological profile; a few credible, specific details are usually enough to convince a target that the person contacting them couldn’t possibly be a stranger. In general, creating a digital profile of a potential victim involves a sequence of steps: consolidating databases, cleaning the data, identifying unique identifiers, extracting content, analyzing behavior, and classifying users.
Maxim Buzinov, an AI product manager at the company “Security Code” (Kod Bezopasnosti), explains that for large-scale, broad-reaching attacks, an AI agent can execute these procedures quite rapidly—often within minutes. In the case of targeted attacks, where it is crucial for the perpetrator to remain undetected, each stage requires additional validation to check for “hallucinations” or erroneous conclusions. This level of analytical work can take anywhere from several hours to several days, the expert points out.
However, even a seemingly simple task—extracting named entities (such as full names or addresses) from text—remains a challenge for machines. Furthermore, the presence of conflicting information across different sources often leads to even greater inaccuracies. For instance, an individual’s place of employment or marital status might appear differently across various sources of leaked data if those sources have not been updated. In such cases, a profile may contain numerous errors—necessitating advanced behavioral verification—or include multiple conflicting versions of the same information. Conversely, if the data points do not contradict one another, this variability actually increases the likelihood of a successful attack. For example, the more passwords and credentials gathered from disparate sources, the greater the chance that at least some will prove effective during an attack. Fraudsters often “delegate” the task of automated credential testing to AI.
“The use of AI agents for compiling digital profiles is still gaining traction, so there isn’t enough data yet to provide a definitive average estimate. Depending on the AI model used and the complexity of extracting the data, the error rate can range from 3–5% for straightforward digital information to 20–30% for data that is more difficult to retrieve,” concludes Buzinov.
About the Author
