I don't want to see the peer review process (which can be weak enough as it is) overwhelmed by a flood of slop. I don't want reviewers to have to suffer reading AI's bad writing either.
Severities are not directly comparable. The CVSS docs leave the choice of the "vulnerable component" as an exercise for the reader. The choice of the "vulnerable component" is also not encoded in the score so a critical may be a critical in the context of a library scored against a theoretical (and bad) program or it could be scored on a single website with known mitigations.
Is the point that you're making that it only found the vulnerabilities because they were in the training set, and therefore AI didn't really "find" them?
If so, I fear you're making the opposite point, which is that AI made genuinely new discoveries that were not in its training set, about codebases that were in its training set. It quite literally expanded the horizon of what we know.
reply