ChatGPT didn’t write this article – I did. Nor did I ask it to answer the question from the title – I will. But I guess that’s just what ChatGPT might say. Luckily, there are some grammar errors left to prove I’m not a robot. But that’s just the kind of thing ChatGPT might do too in order to seem real.
This current robot hipster tech is a fancy autoresponder that is good enough to produce homework answers, research papers, legal responses, medical diagnoses, and a host of other things that have passed the “smell test” when treated as if they are the work of human actors. But will it add meaningfully to the hundreds of thousands of malware samples we see and process daily, or be an easily spotted fake?
In a machine-on-machine duel that the technorati have been lusting after for years, ChatGPT appears a little “too good” not to be seen as a serious contender that might jam up the opposing machinery. With both the attacker and defender using the latest machine learning (ML) models, this had to happen.
Except, to build good antimalware machinery, it’s not just robot-on-robot. Some human intervention has always been required: we determined this many years ago, to the chagrin of the ML-only purveyors who enter the marketing fray – all while insisting on muddying the waters by referring to their ML-only products as using “AI”.
While ML models have been used for coarse triage front ends through to more complex analysis, they fall short of being a big red “kill malware” button. Malware just isn’t that simple.
But to be sure, I’ve tapped some of ESET’s own ML gurus and asked:
A. We are not really close to “full AI-generated malware”, though ChatGPT is quite good at code suggestion, generating code examples and snippets, debugging, and optimizing code, and even automating documentation.
A. We don’t know how good it is at obfuscation. Some of the examples relate to scripting languages like Python. But we saw ChatGPT “reversing” the meaning of disassembled code connected to IDA Pro, which is interesting. All in all, it’s probably a handy tool for assisting a programmer, and maybe that’s a first step toward building more full-featured malware, but not yet.
A. ChatGPT is very impressive, considering that it’s a Large Language Model, and its capabilities surprise even the creators of such models. However, currently it’s very shallow, makes errors, creates answers that are closer to hallucinations (i.e., fabricated answers), and isn’t really reliable for anything serious. But it seems to be gaining ground quickly, judging by the swarm of techies poking their toes in the water.
A. For now, we see three likely areas of malicious adoption and use:
If you think phishing looked convincing in the past, just wait. From probing more data sources and mashing them up seamlessly to vomit specifically crafted emails that will be very difficult to detect based on their content, and success rates promise to be better at getting clicks. And you won’t be able to hastily cull them due to sloppy language mistakes; their knowledge of your native language is probably better than yours. Since a wide swath of the nastiest attacks start with someone clicking on a link, expect the related impact to supersize.
Smooth-talking ransomware operators are probably rare, but adding a little ChatGPT shine to the communications could lower the workload of attackers seeming legit during negotiations. This will also mean fewer mistakes that might allow defenders to home in on the true identities and locations of the operators.
With natural language generation getting more, well, natural, nasty scammers will sound like they are from your area and have your best interests in mind. This is one of the first onboarding steps in a confidence scam: sounding more confident by sounding like they’re one of your people.
If all this sounds like it may be way in the future, don’t bet on it. It won’t all happen all at once, but criminals are about to get a lot better. We’ll see if the defense is up to the challenge.