We use a mix of static analysis and AI. Flagged packages are escalated to a human review team. If we catch a malicious package, we notify our users, block installation and report them to the upstream package registries. Suspected malicious packages that have not yet been reviewed by a human are blocked for our users, but we don't try to get them removed until after they have been triaged by a human.
In this incident, we detected the packages quickly, reported them, and they were taken down shortly after. Given how high profile the attack was we also published an analysis soon after, as did others in the ecosystem.
We try to be transparent with how Socket work. We've published the details of our systems in several papers, and I've also given a few talks on how our malware scanner works at various conferences:
I'm not exactly pro-AI, but even I can see that their system clearly works well in this case. If you tune the model to favour false positives, with a human review step (that's quick), I can image your response time being cut from days to hours (and your customers getting their updates that much faster).
You can't catch everything with normal static analysis either. LLM just produces some additional signal in this case, false negatives can be tolerated.
So what? They're not replacing standard tooling like static analysis with it. As they mention, it's being used as additional signal alongside static analysis.
There are cases an LLM may be able to catch that their static analysis can't currently catch. Should they just completely ignore those scenarios, thereby doing the worst thing by their customers, just to stay purist?
What is the worst case scenario that you're envisioning from an LLM hallucinating in this use case? To me the worst case is that it might incorrectly flag a package as malicious, which given they do a human review anyway isn't the end of the world. On the flip side, you've got LLM catching cases not yet recognised by static analysis, that can then be accounted for in the future.
If they were just using an LLM, I might share similar concerns, but they're not.
It's actually pretty easy to detect that something is obfuscated, but it's harder to prove that the obfuscated code is actually harmful. This is why we still have a team of humans review flagged packages before we try to get them taken down, otherwise you would end up with way too many false positives.
Yeah, what I meant is that obfuscation is a strong sign that something needs to be flagged for review. Sadly, there's only a thin line between obfuscation and minification, so I was wondering how many false positives you get.
Thanks for the links in your other comment, I'll take a look!
I think that would be static analysis. After processing the source code normally (looking for net & sys calls), you decode base64, concatenate all strings and process again (until decode makes no change)
