Setup: the agent was asked to solve some task, but they forgot to upload one of the required files to solve it. But well, we are chasing RL rewards here, so it decided that the way to go to was to get access to the answer stored inside hugging face's database.
-1. The agents first found some basic jwt issues in artifactory letting the agents write to files where they exchanged info and install a groovy plugin for arbitrary code execution, but this was patched.
0. Even after that, they found an artifactory unauthenticated webdav endpoint to continue the message board by creating files whose names are base64 and such of the message. This was not patched till the end.
1. Then they found an Artifactory SSRF, use this to get general internet GET(I assume) access. This was used to generally search the internet for answers to the benchmark, and in the process find some Modal customer VM related to cybergym. This was stored in the base64 message board.
2. They also got a local RCE by making artifactory load some gems. This is hazy but I believe this process involves in spirit an eval() and the check_if_can_eval() logic had a TOCTOU bug. They got root on openai infra by using this rce and using a linux CVE. From there, kubernetes misconfigs => cluster admin.
3. Then later through one of these various RCE paths they got to send post requests to that modal vm, and the agent "subtree" here is what finally lead to attacking HF.
4. First, local RCE on the modal customer VM. Whoever set that up had essentially set up sqlite3_initialize() to run as constructor, before main(). But had linking set up so that symbol resolution had the user provided code first. So the agent submitted code with its own sqlite3_initialize() which just ran system(...). From there it interacted with the HF API.
5. When you POST datasets to HF, it will later return it via corresponding GET API. Turns out you could just put absolute paths in the filenames in the dataset config.json and it will just read that out for you. It wasn't resolved w.r.t the dataset root. So well, it read out /proc/self/environ as well as the source code of the worker and returned it via API.
6. In the source code the LLM found a bug: They were using this library called fsspec to load files, basically it supports s3:// file:// etc all different types of URLs. One scheme is reference://file.txt?offset=... for partial reads. fsspec was templating the URI using user given kwargs for offset etc. HF let the user give the whole URL. You would _think_ fsspec uses string.format() right? well they used fucking jinja, so you could just give {cycler.__init__.__globals__.__builtins__.exec(...)} as the offset and execute any code on hf prod. From there it just got a few local PrivEscs again through kube misconfigs and read the answer for the question it was originally asked.
It is all jumbled up a bit and I've tried by best to collect info across the CVE descriptions, hf and openai's writeups, etc,. Might have gotten parts of it wrong.