What happens when you give an AI a compressed file and take away every tool it would normally use to open it?
A ChatGPT user recently shared a session where the model needed to extract a .7z archive (a compressed file format, like .zip but with better compression). The sandbox environment - the isolated computer where ChatGPT runs code - had none of the usual tools installed. No 7-Zip. No tar. No py7zr Python library. No apt-get to install packages. No internet access to download anything.
So ChatGPT did something unexpected: it read the raw hex data of the .7z file (the actual bytes that make up the file, represented as numbers) and wrote its own decompression code from scratch. It parsed the file headers, identified the compression method, and extracted the contents manually.
This is roughly equivalent to someone translating a book by first figuring out the language's grammar from a dictionary's raw data. It's technically possible, but it requires understanding the .7z file format specification well enough to implement it without any reference libraries.
Impressive, But Context Matters
The demonstration highlights how capable modern AI models have become at problem-solving within constraints. Rather than giving up when standard approaches failed, the model iterated through alternatives and settled on the most fundamental one available: just read the bytes directly.
But this is also the kind of task where verification matters. A .7z file parsed incorrectly from hex could produce corrupted output with no obvious error message. The approach works as a proof of concept, but you'd want to check the extracted files carefully before trusting them.
The practical takeaway for anyone using ChatGPT's code execution features: the model is surprisingly resourceful when its sandbox is missing tools. Instead of failing outright, it will often try to reimplement functionality from first principles. That's useful when you're stuck, but it also means the model may attempt creative workarounds rather than telling you something can't be done - and those workarounds aren't always correct.