Comment by baobabKoodaa
10 hours ago
I don't believe unprivileged people/companies have access to any frontier model that will do cyber work.
10 hours ago
I don't believe unprivileged people/companies have access to any frontier model that will do cyber work.
You don’t need the latest frontier to be “safe”. Current open-weight models are what frontier was 6-12 months ago. Run them slower but internally without safeguards 24/7 to find holes in your software?
I had my Claude refuse to do anything interesting with the recent php vulnerability, yet it'll readily identify issues in the code of web apps it whips up for me, so I don't think that's entirely the case. Though that may be a case of me knowing what I'm doing, so I'm able to draw on my experience and ask specific questions (like why doesn't that bit lead to an XSS?) without realizing that someone newer to the field wouldn't know to ask. Yeah, idk.