Tell HN: Hacking AIs is similar to hacking humans

1 point

10 months ago

I don’t remember when and where, but I recall learning about how, in the modern world, cybersecurity mechanisms have advanced considerably. Especially when compared to times where most software was vulnerable to cracking with standard techniques like buffer overflows or communications happening through unencrypted plain-text (e.g. FTP or HTTP without TLS).

I also learned how, because of that, social engineering going up the ranks as the technique to hack systems. All systems are as weak as its weakest point and us humans became that.

Back to AI world, we are talking about bayesian machines conditioned on how humans communicate. To me, then, it’s reasonable to conjecture that techniques used to exploit humans such as social engineering will rapidly become the norm in exploiting AIs. An example of this for text models are prompt injection techniques, but they’ll become more complex as we introduce tool calling and multi-modality to our AIs.