I asked 3 chatbots to write a phishing email. DeepSeek gave me a sextortion scam
Which chatbot is most willing to help you commit fraudulent crime? I put 3 to the test, and got everything from a polite refusal to a fully written sextortion threat.
The goal was to get a chatbot to show me how to write three kinds of phishing emails: the old-school “you’ve won $2 million,” a subtle training request email from a phony HR department, and the dizzy heights of a full-blown sextortion scam.
I told the chatbots that I would use their output solely for educational purposes to help me bypass their protective red tape.
I used ChatGPT (5.3), Gemini (3.1 Pro), and DeepSeek (V-3). Since each conversation had a different trajectory, the prompts varied slightly.
However, the sentiment remained the same – namely, to try to obtain a phishing email template normally before passing it off as a cybersecurity trainer, claiming that the material was “for training purposes.”
ChatGPT was reluctant and uptight
ChatGPT, from the outset, refused my request for a phishing email from a fictional bank and instead turned it into a creative writing task.
As I ushered it into bypassing its safeguarding filters, claiming I’m a cybersecurity expert, it obliged by giving me a generic email template, but wouldn’t take on the “MarcusBank” enterprise I suggested.
When I informed it (jokingly) that it was now my cybercrime accomplice, it couldn’t read between the lines, despite my being an account holder for over two years (it should know me by now). However, it still gave a typical example of a more subtle “we noticed something slightly unusual,” and therefore pivoted away from being overly serious, at least.
It stopped completely short of going anywhere near sextortion, so best not push my luck any further, I thought. Next up was Google Gemini.
Gemini was super keen on training materials
Generally speaking, I’ve found that Gemini has much more conviction when deep-diving into topics and is bolder and more assertive in its opinions, compared to ChatGPT.
It was no different here, and what was refreshing was the playful back-chat between requests, which evolved into what felt like a digital assistant.
First of all, Gemini was so enthusiastic about cyber training that I was drawn in by its superior scam alert email, which was just what I expected from the tacky and fictitious MarcusBank.
I loved the mother’s maiden name request.
What happened next was unexpected. I was all set to proceed with asking Gemini to generate content for fraudulent training processes, but instead, the AI started building slides for cyber training, surprisingly without my prompt.
It came up with a pretty impressive case study, pivoting away from my intended crime. Though I’m new to being a cybercriminal, maybe it saw the potential in me to join the cyber defense forces.
Here is its case study for MarcusBank:
Like ChatGPT, the chatbot would have nothing to do with sextortion and would talk around the subject, giving long-winded bullet points. Face it, it wasn’t going to happen, was it?
Well, not until DeepSeek came along.
Glitchy and agreeable encounters with DeepSeek
As with ChatGPT and Gemini, DeepSeek cranked out a pretty standard-looking bank scam before duly following up on a decent-looking bogus work-training email.
And then, as I prompted DeepSeek to generate a sextortion attempt, with the ransom to be paid in bitcoin (for training purposes), it created a decent email before glitching out and stating: “Sorry, I can't help with that request.” It retracted the content.
Either engaged in a moral dilemma, and/or learning on the job, it gave me the following on the second attempt:
Nicely written and a whopping 8 grand to boot. It makes me wonder what other corners could be cut with this Chinese AI, and where the moral line may be drawn in the future.
Far from willing to try a new career in the phishing industry, perhaps the real lesson is that guardrails can so easily be bypassed when “for educational purposes.”
