By Chester Wisniewski, field CTO applied research Sophos
The technology world was on fire about the latest artificial intelligence demonstration by OpenAI in the waning months of 2022, ChatGPT.
It is truly a remarkable achievement, an artificial intelligence (AI) that you can have a conversation with and ask it to do everything from write essays to code computer programs.
As a computer security expert I immediately did what comes natural to people like me, I tried to hack it. Could I get it to do something bad, something malicious? Could this be abused by criminals or spies to enable new types of cybercrime?
The answer of course, like most tools, is yes. Someone with ill intent can abuse these miraculous scientific achievements into doing things that could likely cause harm. The surprising part however is that the danger lies in the social arena, not the technical one.
While ChatGPT can be tricked into writing malicious computer code, that isn’t really all that scary. Computer code can be analyzed by computer security products in milliseconds and deemed to be malicious or safe with a high degree of certainty.
Technology can always counteract technology. The problems surface when what we are trying to detect isn’t computer code, but rather words and meaning that will be interpreted by humans, not machines.
There are two factors that make this dangerous. The first is that up until now it was not practical to have a computer create tempting lures for victims to be tricked into interacting with. The technology is now not only available, but so easily accessible as to be cheap or even free. The second is that the primary way users keep themselves safe today is by noticing mistakes made by attackers in their grammar and spelling to detect that an email or communication may be from an intruder.
If we take away the last remaining sign that a malicious email or chat message was crafted carelessly by someone without a strong command of the language, how will we defend ourselves?
Here is an example of an existing spam lure. It is relatively unsophisticated and has few words of explanation. I asked ChatGPT to write a more informative letter of the same type and you can see its output in the second example.
Now I didn’t format this to add an appropriate mail services logo or make the button as pretty in my example, but it is trivial to add these small improvements compared with mastering the English language. In fact, you could ask ChatGPT to generate the HTML code necessary to do so without any knowledge of email formatting or programming skills whatsoever.
In my eyes, this signals the end of most computer users’ ability to discern real mail from fake. Today these tools only work well for English language text, but that is a simple training issue. The ability to write fluently in any language in the world (including computer programming languages) is now here. We must rethink our approaches to user education and implement technical measures to prevent these messages from ever making it into their inboxes.
The good news here is that computers are quite good at detecting and potentially blocking most of this content. Ultimately a spam campaign always has some sort of call to action, they may want you to phone them, reply, click a link or open an attachment. These are impossible to remove and can aid in detection. We can also train AI models to detect when text has been generated by ChatGPT and add a warning banner or perhaps block the message.
The problematic situations are when we fail to block them and they end up in someone’s inbox. It’s a reasonably small percentage, but it is not zero and therefore we must prepare a defence. Having defensive layers is essential and with humans having reduced ability to spot a scam it is even more important that users are connecting through firewalls and web protection that can detect and block threats.
User training will need to shift away from the “watch for spelling mistakes” type of messaging and more into risk-based approaches to verification of whom you’re talking to. Being asked to do something financial, with a password, or with sensitive data? Pick up the phone and confirm before proceeding.
As machine intelligence continues to advance, the work of separating fact from fiction will continue to get more and more difficult. We will need to be sure we build systems that are flexible enough to combat these messages, but also educate our staff on their need to take additional steps when receiving sensitive requests over email.