One of the most insidious parts of this malware's payload, which isn't getting enough attention, is how it chooses the replacement wallet address. It doesn't just pick one at random from its list.
It actually calculates the Levenshtein distance between the legitimate address and every address in its own list. It then selects the attacker's address that is visually most similar to the original one.
This is a brilliant piece of social engineering baked right into the code. It's designed to specifically defeat the common security habit of only checking the first and last few characters of an address before confirming a transaction.
I'm a little confused on one of the excerpts from your article.
> Our package-lock.json specified the stable version 1.3.2 or newer, so it installed the latest version 1.3.3
As far as I've always understood, the lockfile always specifies one single, locked version for each dependency, and even provides the URL to the tarball of that version. You can define "x version or newer" in the package.json file, but if it updates to a new patch version it's updating the lockfile with it. The npm docs suggest this is the case as well: https://arc.net/l/quote/cdigautx
And with that, packages usually shouldn't be getting updated in your CI pipeline.
Am I mistaken on how npm(/yarn/pnpm) lockfiles work?
Not the parent, but the default `npm install` / `yarn install` builds will ignore the lock file unless everything can be satisfied, if you want the lock file to be respected you must use `npm ci` / `yarn install --frozen-lockfile`.
In my experience, it's common for CI pipelines to be misconfigured in this way, and for Node developers to misunderstand what the lock file is for.
Welcome to the web side. Everything’s bonkers. Hard-earned software engineering truths get tossed out, because hey, wtf, I’ll just do some stuff and yippee. Feels like everyone’s stuck at year three of software engineering, and every three years the people get swapped out.
That's because they are being "replaced", in a sense!
When an industry doubles every 5 years like web dev was for a long time, that by the mathematical definition means that the average developer has 5 years or less experience. Sure, the old guard eventually get to 10 or 15 years of experience, but they're simply outnumbered by an exponentially growing influx of total neophytes.
Hence the childish attitude and behaviour with everything to do with JavaScript.
Sorry, I had assumed this was what you were doing when I wrote my question but I should have specified. And sorry for now making your npm install step twice as long! ;)
npm ci should be much faster in CI as it can install the exact dependency versions directly from the lockfile rather than having to go through the whole dependency resolution algorithm. In CI environments you don't have to wait to delete a potentially large pre-existing node_modules directory since you should be starting fresh each time anyway.
Yeah, I think I had made the assumption that they were using `npm ci` / `yarn install --frozen-lockfile` / `pnpm install --frozen-lockfile` in CI because that's technically what you're always supposed to do in CI, but I shouldn't have made that assumption.
As others have noted, npm install can/will change your lockfile as it installs, and one caveat for the clean-install command they provide is that it is SLOW, since it deletes the entire node_modules directory. Lots of people have complained but they have done nothing: https://github.com/npm/cli/issues/564
The npm team eventually seemed to settle on requiring someone to bring an RFC for this improvment, and the RFC someone did create I think has sat neglected in a corner ever since.
Is there no flag to opt out of this behavior? For Rust, Cargo commands will also do this by default, but they also have `--offline` for not checking online for new versions, `--locked` to require sticking with the exact version of the lockfile even when allowing downloading dependencies online (e.g. if you're building on a machine that's never downloaded dependencies before, so they aren't cached locally, but you still don't want to allow implicit updates), and `--frozen` (which is a shorthand for both `--locked` and `--offline`). I'm honestly on the fence about whether this is even sufficient, since I've worked at multiple places where the CI didn't actually run with `--locked` because whoever configured it didn't realize, and at least once a surprise update to the lockfile in CI ended up causing an issue that took a bit of time to debug before someone realized what was going on.
You’re right and the excerpt you quoted was poorly worded and confusing. A lockfile is designed to do exactly what you said.
The package.json locked the file to ^1.3.2. If a newer version exists online that still satisfies the range in package.json (like 1.3.3 for ^1.3.2), npm install will often fetch that newer version and update your package-lock.json file automatically.
That’s how I understand it / that’s my current knowledge. Maybe there is someone here who can confirm/deny that. That would be great!
We should be displaying hashes in a color scheme determined by the hash (foreground/background colors for each character determined by a hash of the hash, salted by that character's index, adjusted to ensure sufficient contrast).
That way it's much harder to make one hash look like another.
As someone with red/green vision deficiency: if you do this, please don’t forget people like me are unable to distinguish many shades of colours, which would be very disadvantageous here!
I think 9dev was saying that providing only a colorized version might make it unreadable to some people, not merely that they wouldn't benefit from the extra color information.
There's actually nothing the developers can do about this particular issue other than to display all colors and allow colorblind people to see the colors that they can see.
It doesn't matter which colors the algorithm chooses so long as background/foreground are very distinguishable to as wide an audience as possible, and prev/next are likely to be distinguishable more often than not.
That's a lot of flexibility within which to do clever color math which accounts for the types of colorblindness according to their prevalence.
For the newly made up feature, which doesn't exist yet, but already has an issue?
Simple. Instead of forcing colour, one could retain a no colour option maybe?
Done. Solved.
Everything should have this option. I personally have no colour vision issues, other than I find colour annoying in any output. There's a lot who prefer this too.
Agreed, although I would argue that maximal hash contrast should be default, and if people find they prefer less, they can turn it down.
If you're the sort of person who would think about adjusting it to suit your sensitivity to this kind of attack, you're likely not the sort of person that the feature is trying to protect anyhow.
Not sure why you're being downvoted, OpenSSH implemented randomart which gives you a little ascii "picture" of your key to make it easier for humans to validate. I have no idea if your scheme for producing keyart would work but it sounds like it would make a color "barcode".
I have to say the openssh random art has never really helped for me - I see each individual example so infrequently and there's so little detail to remember that it may as well just be a hash for all the memorability it doesn't add
If you ignored the characters and just focused on the background colors, yeah I suppose it would look like a barcode. But the way I envision it, each line on the barcode is a character, so it still copy/pastes into notepad as the original text, but it'll copy/paste into word as colored text with colored background.
> This is a brilliant piece of social engineering baked right into the code. It's designed to specifically defeat the common security habit ...
I don't agree that the exuberance over the brilliance of this attack is warranted if you give this a moment's thought. The web has been fighting lookalike attacks for decades. This is just a more dynamic version of the same.
To be honest, this whole post has the ring of AI writing, not careful analysis.
I've been thinking about using Levenshtein to make hexadecimal strings look more similar. Levenshtein might be useful for correcting typos, but not so when comparing hashes (specifically the start or end sections of it). Kinda odd.
It actually calculates the Levenshtein distance between the legitimate address and every address in its own list. It then selects the attacker's address that is visually most similar to the original one.
This is a brilliant piece of social engineering baked right into the code. It's designed to specifically defeat the common security habit of only checking the first and last few characters of an address before confirming a transaction.
We did a full deobfuscation of the payload and analyzed this specific function. Wrote up the details here for anyone interested: https://jdstaerk.substack.com/p/we-just-found-malicious-code...
Stay safe!