ChatGPT Capabilities: Cryptography Challenges (Part I)
We haven’t published anything related to ChatGPT until now, mostly due to the fact that our use of GPT has been constrained to offloading menial tasks that are repetitive or just needlessly tedious yet simple enough. The state of the art as far as GPT is involved has changed swiftly over the course of 2023 and 2024, at times providing very unreliable results.
The model has some clear biases and the pitfalls are well known, the most relevant of them being the fact that “it does not know what it does not know”, which, applied to topics like programming related prompts, means it often produces strange concoctions: code where outdated APIs might be used mixed with newer ones, insecure code constructs (far more than people seem to notice or assume), poorly organized code (for example, GPT-4 produced less than stellar Python source code in response to relatively simple prompts), etc.