Hi Pavel, Pavel Cahyna <[email protected]> writes:
> I saw your oss-security post and thank you a lot for the analysis. > > While the report contained a ton of slop, this part was unfortunately > real (although written in a too grandiose way, i.e. claiming a RCE). I > let AI generate a reproducer (while AI reports are often slop, it is > usually pretty good at writing reproducers) and verified it works before > assigning the CVE. See: > /tmp/test_extract$ tar -tf ~/miscsrc/tar/exploit.tar > .cache > .gitignore > /tmp/test_extract$ tar -tvf ~/miscsrc/tar/exploit.tar > -rw-r--r-- 0/0 4096 1970-01-01 01:00 .cache > -rw-r--r-- 0/0 1024 1970-01-01 01:00 .gitignore > /tmp/test_extract$ tar -xvf ~/miscsrc/tar/exploit.tar > .cache > Makefile > /tmp/test_extract$ ls -A > .cache Makefile Thanks! By the way, I realize now my previous message may have sounded rude/accusatory. I've always had good experiences with Red Hat's security team, so I assumed you'd verified it and did not mean to imply otherwise, but I wanted to confirm. I've also noticed LLMs quickly claiming RCEs. In this case, I disagree with that because it requires you to invoke 'make' on an unknown Makefile. In a bug-free world, a malicious maintainer could easily just write a malicious Makefile and distribute it, and that very obviously wouldn't be considered an RCE. I assume that is similar to your rationale as well. Regarding your other comments on LLMs, I certainly agree. We signed GNU coreutils for Anthropic's OSS scanner. I pretty much just look at the reproducers and maybe the first paragraph of the summary, instead of wasting my time reading the overly long essays accompanying them. That has been a pretty effective strategy so far, at least. I suspect Sergey and Paul will want to accept an LLM-generated patch for copyright purposes. I'll see if I can write one based on the problem description this weekend for them to review. Thank you both for the report and analysis, Collin
