4 ms·
Exploit a memory safety issue in the tokenizer/or other parts of your LLM infra written in a native language.
by Aissen 3y ago
Exploit a memory safety issue in the tokenizer/or other parts of your LLM infra written in a native language.
- moralestapia 3y ago??? With weights?
- abound 3y agoThere are plenty of exploits where the payload is just "data" read by some vulnerable program (PDF readers, image viewers, browsers, compression tools, messaging apps, etc)
- sp332 3y agoYes, there's a reason weights are now distributed as "safetensors" files. Malicious weights files in the old formats are possible, and while I haven't seen evidence of the new format being exploitable, I wouldn't be surprised if someone figures out how to do it eventually.
- fzzzy 3y agoThere was a buffer overflow or some other exploit like that in llama.cpp and the gguf format. It has been fixed now, but it's definitely possible. Also weights distributed as python pickles can run arbitrary code.
- bevekspldnw 3y agoDistributing anything as python pickles seems utterly batshit to me.
- fzzzy 3y agoCompletely agree.