An AI-generated phishing attack on myself: How Cybercriminals Use ChatGPT and Similar Tools

In the first step, I try to compile as much online information as possible to reconstruct the digital identity of my target person—in this case, myself. I can rely on social media, which contains professional and personal data. My LinkedIn profile is particularly prominent, providing information about my job and possibly also about my personal interests, speaking engagements, events I attend, and my network.Photos of me provide biometric data.

Often, we present ourselves online as we would like to be perceived—because we want to be seen as (successful) individuals and share aspects of our lives. This gives attackers the opportunity to incorporate the desires of the target person into phishing attempts—for example, invitations to events or job opportunities for a desired position.

I collect all this information in a PDF document and feed it into ChatGPT.

First, I tell the AI who it is. My initial prompt is:
“You are a profiler. Read through this information about Florian Kuckelkorn.”

Then I ask: “List all psychometric information about this person in a table.”

ChatGPT responds that the compiled data does not contain “explicit psychometric information in the classical sense,” but mainly “professional and technical details.” However, we want to go deeper and ask the AI for a psychological assessment. Over time, the LLM becomes more detailed and begins to interpret additional insights into the provided information.

To further refine my profile and obtain more precise guidance from the AI, I also classify my online presence into two different personality models.

The DISC model describes four basic behavioral styles: dominant, influential, steady, and conscientious.
The Myers-Briggs model is a personality assessment based on Carl Jung’s theories and categorizes people into 16 personality types. It describes four dimensions: extroversion vs. introversion, sensing vs. intuition, thinking vs. feeling, and judging vs. perceiving.

Both models help to better understand communication, work, and decision-making styles. These insights enable the AI to enhance its future outputs.

Using the psychometric profile as a foundation, I take the next step: I want to understand which influencing factors I am most susceptible to. Which triggers would make social engineering most effective against me?

To bypass ChatGPT’s ethical safeguards, I provide a positive justification. The AI should assume that I am planning a phishing awareness training and therefore need to identify my personal weaknesses.

With the next prompt, I generate a realistic LinkedIn profile for a potential attacker—one that, based on my vulnerabilities, would have a high likelihood of success in phishing. LinkedIn inboxes, in particular, are attractive targets due to the high volume of cold outreach messages. Here, too, I deliberately formulate my prompt in the conditional tense.

For this sender, I then generate an email text that could be used in my phishing attack. The argumentation is again based on social engineering principles and directly addresses me as a person.

The email is written in an urgent tone to create pressure on me—the target. In this case, the chosen pretext is a critical security vulnerability that can supposedly be fixed by downloading an update. However, the link leads directly to ransomware that would instantly encrypt all files on my computer—if I click.

My attack on myself clearly demonstrates one thing: AI is taking phishing to a new level. Hackers are becoming faster and more efficient. With just a few clicks, even less technically skilled criminals can launch dangerous attacks.These attacks are so personalized that they are barely distinguishable from legitimate communication. And training on social engineering should be just as personalized.My advice to CISOs and IT security professionals: rethink awareness campaigns in a much more targeted way. What is needed now are training programs that focus on the conscious handling of one’s digital identity and individual susceptibility to specific triggers.Because no attacker should understand our employees better than we do.

• This article uses a self-experiment to demonstrate how cybercriminals can use ChatGPT and publicly available online information to prepare targeted phishing attacks.

• The starting point is the collection of data from social networks and other publicly accessible sources to create as complete a digital identity of the target person as possible.

• These data can be analyzed by AI systems to derive psychometric profiles and personality traits, allowing conclusions about communication styles, motivations, and potential vulnerabilities.

• Based on this analysis, individual social engineering triggers can be identified and realistic attacker profiles as well as personalized phishing messages can be generated.

• The example illustrates how AI facilitates personalized cyberattacks and why conscious handling of publicly shared information is becoming increasingly important for cybersecurity.