I’m trying to feel more comfortable using random GitHub projects, basically.
I don’t think “AI” is going to add anything (positive) to such a use case. And if you remove “AI” as a requirement, you’ll probably get more promising candidates than if you restrict yourself to “AI” (whatever that means) solutions.
I don’t care if the solution is AI based or not, indeed.
I guess I thought it like that because AI is quite fit for the task of understanding what might be the purpose of code in a few seconds/minutes without you having to review it. I don’t know how some non-AI tool could be better for such task.
Edit: so many people against the idea. Have you guys used GitHub Copilot? It understands the context of your repo to help you write the next thing… Right? Well, what if you apply the same idea to simply review for malicious/unexpected behaviour on third party repos? Doesn’t seem too weird for me.
AI is quite fit for the task
EXTREMELY LOUD INCORRECT BUZZER
AI is quite fit for the task of understanding what might be the purpose of code
Disagree.
I don’t know how some non-AI tool could be better for such task.
ClamAV has been filling a somewhat similar use case for a long time, and I don’t think I’ve ever heard anyone call it “AI”.
I guess bayesian filters like email providers use to filter spam could be considered “AI” (though old-school AI, not the kind of stuff that’s such a bubble now) and may possibly be applicable to your use case.
Bayesian filters are statistical, they have nothing to do with machine learning.
If you’re talking about naive bayes filtering, it most definitely is an ML model. Modern spam filters use more complex ML models (or at least I know Yahoo Mail used to ~15 years ago, because I saw a lecture where John Langford talked a little bit about it). Statistical ML is an “AI” field. Stuff like anomaly detection are also usually ML models.
The A* algorithm doesn’t have anything to do with machine learning either, but the first time I ever learned about it was in a computer science class in college called something like “Introduction To Artificial Intelligence”.
But it’s very much the case that the term “AI” has a very different meaning now-a-days during this cringy bubble than it did back in 2004 or 2005 or whenever that was.
Today “AI” is basically synonymous with “BS”. Lol.
AI is quite fit for the task of understanding
Sure, and parrots are amazing at spotting fallacies like cherry picking…
Don’t listen to the idiots downvoting you. This is absolutely a good task for AI. I suspect current AI isn’t quite clever enough to detect this sort of thing reliably unless it is very blatant malicious code, but a lot of malicious code is fairly blatant if you have the time to actually read an entire codebase in detail, which of course AI can do and humans can’t.
For example the extra
.that disabled a test in xz? I think current AI would easily be capable of highlighting it as wrong. It probably wouldn’t be able to figure out that it was malicious rather than a mistake yet though.I mean anything is a good fit for future, science fiction AI if we imagine hard enough.
What you describe as “blatant malicious code” is probably only things like very specific C&C domains or instruction sets. We already have very efficient string matching tools for those, though, and they don’t burn power at an atrocious rate.
You’ve given us an example so PoC||GTFO. Major code AI tools like Copilot struggle to explain test files with a variety of styles, skips, and comments, so I think you have your work cut out for you.
We already have very efficient string matching tools for those, though
How is a string matching tool going to find a single
.?You’ve given us an example so PoC||GTFO
🙄
A single character, per your definition, is not blatant malicious code. Stop moving the goalposts.
It’s clear you don’t understand the space and you don’t seem to have any interest in acting in good faith based on your other comments so good luck.
I’m not moving any goalposts. The addition of the
.was very blatant. They literally just added a syntax error. It went undetected because humans don’t have the stamina to exhaustively do code review down to that level. Computers (even AI) don’t have that issue.You are clearly out of your depth here.
Privado CLI will produce a list of data exfilration points in the code.
If the JSON output file points out a bunch of endpoints you don’t recognize from the README, then I wouldn’t trust the project.
Privado likely won’t catch a malicious binary file, but your local PC antivirus likely will.
The solution to what you want is not to analyze the code projects automagically, but rather to run them in a container/virtual machine. Running them in an environment which restricts what they can access limits the harm an intentional — or accidental bug can do.
There is no way to automatically analyze code for malice, or bugs with 100% reliability.
Of course, 100% reliability is impossible even with human reviewers. I just want a tool that gives me at least something, cause I don’t have the time or knowledge to review a full repo before executing it on my machine.
That is another tool you can use to reduce the risk of malicious code, but it isn’t perfect, so using sandboxing doesn’t mean you can forget about all other security tools.
There is no way to automatically analyze code for malice, or bugs with 100% reliability.
He wasn’t asking for 100% reliability. 100% and 0% are not the only possibilities.
What do you consider malicious, specifically. Because AI are not magic boxes, they are regurgitation machines prone to hallucinations. You need to train it on examples to identify what you want from it.
I just want a report that says “we detected in line 27 or file X, a particular behavior that feels weird as it tries to upload your environment variables into some unexpected URL”.
particular behavior that feels weird
Yea, AI doesn’t do feelings.
tries to upload your environment variables into some unexpected URL
Most of the time that is obfuscated and can’t be detected as part of a code review. It only shows up in dynamic analysis.
AI doesn’t do feelings
How can I have a serious conversation with these annoying answers? Come on, you know what I am talking about. Even an AI chatbot would know what I mean.
Any AI chatbot, even “general purpose” ones will read your code and will return a description of what it does if you ask it.
And particularly AI would be great at catching “useless”, “weird” or unexplainable code in a repository. Maybe not with the current levels of context. But that’s what I want to know, if these tools (or anything similar) exist yet.
Thank you.
Questions about AI seem to always bring out these naysayers. I can only assume they feel threatened? You see the same tedious fallacies again and again:
- AI can’t “think” (using some arbitrary and unstated definition of the word “think” that just so happens to exclude AI by definition).
- They’re stochastic parrots and can only reproduce things they’ve seen in their training set (despite copious evidence to the contrary).
- They’re just “next word predictors” so they fundamentally are incapable of doing X (where X is a thing they have already done).
AI doesn’t do feelings
It absolutely does. I don’t know where you got that weird idea.
Honey your AI girlfriend doesn’t actually love you
Define love. Good luck.
You’re right, I hope the two of you are very happy
This absolutely sent me.
Not exactly what you asked, but related; roast your Github profile: https://github-roast.pages.dev/
How is that related? I don’t see it.
It’s an AI tool analyzing a Git repo.
It doesn’t analyze only one repo
Perhaps snyk.io I used it in the past, but I didn’t find it quite useful. Now I have a github action to upgrade dependencies every week. But you want some kind of scanner to be more involved on the actual codebase. Did you look into https://github.com/marketplace?query=security ? That’s what I would do. But I never heard of any of those listed there. Let us know your findings after some time if you test 'em ;) good luck!
