yeah but they can just put the dropper, etc in index.js, so that it runs at import time rather than at install time, no? i guess first-install time is often a privileged developer machine, and will execute in a "server"-like runtime such as Node, Bun, Deno. but blocking preinstall scripts is basic first aid...
Nobody is claiming this is a complete solution to security. I would call this "necessary but not sufficient". You won't get far as an engineer if you refuse to implement "necessary but not sufficient" changes because the change doesn't in and of itself one-shot the entire problem.
These half-measures is why everything has gone to shits. Instead of properly auditing software, reducing the quantity and increasing the quality, we keep pushing more and more garbage where all you find is 2FA that, captcha this, not supported this, app not signed, etc..
Can't agree more. I'm working on OSS security tool that can protect you from Shai-hulud, no half measures.
The idea is to have a local proxy that injects real secrets into requests in-flight, so a compromised library has nothing to steal because it never had access to any of your secrets: https://github.com/inflightsec/agent-vault-proxy
Defense in depth is the “meat and potatoes” of security.
In other words: people should be auditing their software, but we should also design systems and schemes that provide varying degrees of defense and protection when people invariably fail to review the code they run.
Defence in depth is not just 'throw anything in that might improve security' though. The idea is to have multiple strong layers, not a hundred half-measures that are all easily bypassed. A stronger layer might be sandboxing, or separating your build and publishing steps as others have suggested (and also probably worth restricting the credentials the publishing step to just the relevant packages as well). These will at least robustly prevent a malicious dependency from spreading horizontally, but you'll still potentially ship malware to your customers.
>The idea is to have multiple strong layers, not a hundred half-measures that are all easily bypassed
Definitely. On the other hand, in my opinion, "not running arbitrary code during package install" is not a "half-measure", it's a basic sanity. This whole arbitrary code execution at install time is a convenience feature that was adapted by some package managers, but it was never a good idea.
Fortunately, nixos solves that for me in most cases.
doesn't nixos still have execution hooks when you actually activate a package to use it? Arguably it's just separating the download and install steps (which I do agree, every package manager should have a 'safely download this package so I can inspect it' function as well as a 'make this package ready to use' function, whatever you call those operations. I think 'pip download', for example, managed to violate that expectation which is not particularly sensible)
The fact that it's not sufficient also largely means it's not necessary either, because the real solution is auditing and trusting the codebase as a whole. All you do when disabling install hooks is make a lot of situations much more difficult to handle.
You will also not get very far in engineering if you have a production incident, someone proposes a thing that will mitigate it partially but significantly, and you insist that we can't deploy that mitigation because we need to do the multi-year project that will actually fix it instead. Even if we still need that project, we also need that mitigation.
Sure, solve the problem of "auditing and trusting codebases". Go for it. Shouldn't take you until, oh, let me be generous and give you until next Monday. No sweat.
Feel free to turn it off yourself, just don't be surprised when the attackers switch tactics. And don't make life harder for everyone else by pushing it as mandatory.
(if you want a better mitigation: don't automatically update dependencies in CI. At a minimum have a cooling-off period that you only bypass on manual review. This is not hard to implement and at least gives some time for alarm bells to be sounded before you're pwned. Probably while doing the updates you can also just do a quick sanity check on the changes, that'll also catch the ones using a blatant install hook)
I mostly use languages where the feature doesn't even exist. We don't generally miss it.
I also blocked a proposed mandatory dependabot at work a few months back because I didn't like the way it created a pipeline for any hacker to push a hacked dependency straight into someone's codebase. I'm lifting my objection now that dependabot is defaulting to a 3-day cooldown, though the code bases I'm managing I'm setting to 7 days. (Not to be behind everyone else; I'd be fine if everyone joined me at 7 days. I don't really accept the freeloading objection, there's plenty of entities scanning things now no matter what cooldown you set. I just think 3 is a little tight to expect the full discovery and remediation to take place.)
So, whatever sort of "but it's really hard! I bet you don't do it yourself" implications you may be trying to draw fall very flat. Or whatever you may have been trying to imply about trusting this step but not trusting others... no, I do defense in depth. Giving up on defense in depth because one step isn't enough... well, I think I've probably played the "you won't get very far in engineering" card enough, but hey, here's one last time.
Out of curiosity, which languages? The equivalent operations for C and C++ do generally involve a lot of arbitrary code execution to do the build, for example, they just don't have one central package manager and repository. I can't think of any package managers off the top of my head that don't have some similar feature, but I am obviously not familiar with all of them.
You can see my other responses in this thread about what I think of this in the context of defence in depth. I think disabling install hooks is a very weak defence with substantial downsides and defence in depth does not mean just taking every option that might foil some attack. There are much better means to get a robust extra layer of defence in this context, which is mainly about protecting your publishing pipelines from become a vector for a worm.
> if you want a better mitigation: don't automatically update dependencies in CI. At a minimum have a cooling-off period that you only bypass on manual review. This is not hard to implement and at least gives some time for alarm bells to be sounded before you're pwned.
One of the problems a lot of companies are having right now is trying to determine a proper window size between CVE release and patching.
On one hand, you have LLM created 0days and attacks are happening almost as fast as CVEs can be posted, and CVEs are being posted at lightning speed. You maybe had a month or a few weeks to patch before, now the window may be down to days or even hours.
On the other hand, you want to prevent these repo bombing attacks.
There's a tension there but I think we're going to end up measuring that window in hours within the year, if we're not already there.
If you need quick responses to such things, you're gonna need some intelligence in the loop regardless, even if it's just deciding when it's worth pushing a new release to prod.
Always has been, arguably. Either way I don't think most people have their CI configured to automatically re-run when a dependency updates (imagine the thundering herd!), the compromises are more incidental to runs happening for other reasons, and that configuration neither guarantees a quick response nor protects against supply chain attacks.
(If you find you do need updates when you're sleeping, you probably need to delegate that process or at least the process of checking the update to someone you trust, and you're probably paying for the service. This blog post is in large part an ad for just such a service).
This is moving the goalposts. The original problem was that installing a library should not execute code from that library.
In most sane environments, like for example native languages, this is already the case. Downloading a .dll file and putting it in an appropriate directory won't, by itself, execute code in that library.
You may argue that the code will get executed at some point anyway, but that's besides the point. Sandboxing the build environment is a different problem than sandboxing the test/staging/production environment.
I think we both agree that "adding random obstacles that don't actually protect anything" is not a valid approach to security, but my mental model of the build step is "transformation of input data into output data", and while this step may produce a malicious output from malicious inputs, it should not do anything malicious itself. For example, "gcc source.c" should not execute arbitrary code by itself.
>In most sane environments, like for example native languages, this is already the case.
Installing a node package is much more like compiling a dll, not downloading it. The same is true for most package managers that exist for C and C++ as languages as opposed to for an OS. These are two different tools for different use-cases. (though still pretty much all installation processes for all OSs involve an opportunity for arbitrary code execution, as well, apart from just downloading a zip file and extracting it, which is not the norm)
Neither is 'close the gate with no fence on either side of it'. If you want to run code, you either need to run it in a sandbox or trust it. Choosing to run only part of the code is not really a solution.
(if you want to disable such hooks yourself, then you may get some security by diversity because you're not using the common configuration. But if it becomes the default then these worms will switch to a different vector)
Prod tends to have less privileges than CI/CD. CI/CD tends to be full admin, so it's far more sensitive. Prod tends to have tooling for detecting breaches, better logging, etc. People tend to use containers, which act as a sandbox.
Prod also won't be wormable the way that CI/CD is. With CI/CD I can own another dev, use their creds to push another malicious build script, etc. "Attacker is in my prod env" isn't wormable.
Yes, capabilities in prod would be hugely beneficial but removing CI/CD is massive as a win.
Dev laptops tend to have better monitoring than CI/CD so I still think this is a better option. You can also have devs use VMs or separate dev environments like an ec2 instance.
To be clear, just solving the CI/CD portion is insufficient, but it is a massive win.
iirc does pnpm not allow them by default. But even if we killed them off there would still be a chance of the malware hooking into something else or only working in cli applications.
You'll just end up with people running `./configure` scripts or whatever instead. The solution I've currently landed on is:
1. Audit build scripts/ proc macros for rust code (and mark with cargo-vet).
2. Have an isolated workflow for "build/test/push artifact to temporary place" (s3, github artifact, whatever). No API keys in this workflow.
3. Have another workflow that has the API keys to publish that grabs the artifact and then places it into a registry.
This creates clear separation of "code runs here" and "environment has privileges".
In my own slop-driven programming language I have build scripts declare their capabilities upfront so that you can statically reason about them (same with runtime permissions).
It sucks that we have this glass-jaw dependency system, which is really the main reason these supply chain attacks work.
Really hard to clean up, too. These days, you (being the blackhat) would likely send agents to leverage every compromised repo/app/Web site, almost the instant it comes online, so even if the original mess is cleaned up, there's still a ton of knock-on compromises.
Does anyone happen to have a grep or similar that helps me check if this is showing up anywhere in the trillions of files in node_modules (or pnpm store)?
The what happened section mentions the addition of the `setup.mjs` and `Math_Symbol.js`, along with a change in `package.json` to add `"preinstall": "node setup.mjs"`, so presumably checking for any of those would be a good indication to check further.
can you search all installed node modules for any sign of the shai hulud supply chain attack? What happened
Every package in the family received two new files, setup.mjs and Math_Symbol.js, along with a "preinstall": "node setup.mjs" entry added to each package.json. Anyone who ran npm install against an affected version would have had setup.mjs execute automatically before their install completed.
setup.mjs is a heavily obfuscated dropper. Its only job is to silently download the Bun JavaScript runtime from github[.]com/oven-sh/bun/releases/download/bun-v1.3.13/ and use it to execute the real payload, Math_Symbol.js:
execFileSync(<bun binary>, ['<script_dir>/Math_Symbol.js'], {
stdio: 'inherit',
cwd: <script_dir>
})
The Math_Symbol.js is a heavily obfuscated 728 KB JavaScript file containing credential stealers that harvest secrets from the victim's environment, encrypt the findings, and exfiltrate them to a public GitHub repository whose description reads "Shai-Hulud: Here We Go Again". The payload also contains worm-like propagation functionality to infect packages of other maintainers that have installed one of the compromised packages.
Packj uses static+dynamic code/behavioral analysis to scan for indicators of compromise (e.g., spawning of shell, use of SSH keys, network communication, use of decode+eval, etc). It also checks for several metadata attributes to detect impersonating packages (typo squatting).
It's time for all developers to learn about devcontainers and use them consistently. They're super easy to setup and run and would protect from most of what this worm does. https://code.visualstudio.com/docs/devcontainers/containers is the best guide to get started if you use vscode.
They help only to the extent that you have completely isolated credentials: the hard part isn’t the container, it’s things like fastidiously using separate least-privilege credentials everywhere and not using tools or editor integrations which don’t support that style of work. Once you map your GitHub or AWS token into a container, it’s no longer useful as a security boundary.
Once again, I ask myself: should we start "shaming" developers who don't use isolation? It still seems I am the exception and most people run their dev environment with full permissions. Why?
Yes, it gets boring that each time one of these supply chain attack article appears everybody starts talking about cooldowns, 2FA, MFA, etc. Just don't give Node the permissions to your complete filesystem / network.
I think it’s premature before a lot of tools improve to make that more workable: for example, if you use AWS how realistic is maintain separate IAM for each tool you run and map the right one into a sandbox for each tool? To use your editor’s GitHub integration with a token which can do basic operations and only retrieves a high-privilege token with a hardware presence check when you cut a release?
Theoretically you can do it but the friction is enough to make it non-viable.
Could you not use something like https://github.com/superfly/tokenizer for AWS? They list it as an explicit example, but I have little experience with AWS.
Possibly, but my point was basically that I wouldn’t be judgey about developers not doing something which most tools aren’t designed to make easy, or even possible.
We absolutely should be trying to get to the point where it’s easy - this is like software deployment before containers and shouldn’t be.
Couldn't GitHub detect a Shai-Hulud exfil repo being created and just... block it? Given that it's a worm, the attacker wouldn't be able to adapt all that quickly.
We're talking about Microsoft, who created a notepad.exe that can run an RCE with an LLM bypass prompt.
In the previous Miasma waves, Microsoft was so overwhelmed that they delayed the VSCode extension installs for a couple days with a timeout; literally not understanding what was going on and neither how the malware was spreading.
microsoft those dudes are hilarious this decade plus old github request is the only one i've ever gotten email alerts for it just made me lmao more each and every year. someone dropped a cake on its tenth birthday https://github.com/microsoft/vscode/issues/519#issuecomment-...
I am kind of surprised GitHub doesn't seem to have built a simple classifier for public repos to proactively lock the account of anyone uploading such obviously fishy things (for their own good, at least before the repo is publicly findable), so it can't be used as a rendezvous.
Surely Github's software is good enough that an intern can slop the 80/20 together in a day? It would be an actually good use of AI spending.
> I am kind of surprised GitHub doesn't seem to have built a simple classifier for public repos to proactively lock the account of anyone uploading such obviously fishy things (for their own good, at least before the repo is publicly findable), so it can't be used as a rendezvous.
Maybe that's the idea. Regards, the <insert your favourite 3 letter agency here>
"Update — August 4, 2026, 13:37 CEST: At least 434 packages (across 1381 versions) have been compromised by the worm, with a combined total of over 2 billion monthly installs at the time of writing."
You know with all this AGI swirling around nowadays that is stronger than nation state hackers you think one of these companies would demonstrate just how capable they are by defending public infrastructure.
what commercial tools are enterprises using today to defend against such attacks? Do they really work? I mean, do they report/block malware after the fact or detect proactively. Because if the latter then, package registries should really be removing reported packages, right?
Here's the updated Antimiasma tool for mitigation [1] [2]. More details on how this tool was built and how the Miasma worm works on my website [3].
This is the first false flag in the campaign series, where setting the "LANG" environment variable to "ru_RU.UTF-8" or "ru_RU.KOI8-R" won't stop the spreading mechanism.
So it's likely this could've been any script kiddie that modified the TeamPCP source code dump. It could now also be still APT28/29, that was kind of the purpose of the code dump... to gain plausible deniability :)
This is the most boring comment posted on every one of these. NPM is by no means the worst offender here and offers a ton of solutions to this, lots of package managers are behind it or equivalent.
NPM gets targeted a lot because it's popular. That's it.
It’s correct that NPM is not unique but it is the worst for cultural reasons: no other ecosystem started with such a limited language, which lead to the culture of publishing tons of small packages working around things which everything else had builtin. A Python project which has a hundred dependencies is considered quite large but the median React project had north of 30 thousand for years and years.
It’s not that there’s a single stdlib feature which would’ve stopped this but more that JavaScript developers have been conditioned that it’s normal to install tons of packages and update them quite frequently so there are a lot of individual maintainers who if compromised have a surprising impact.
You’re exposed as a function of the number of dependencies so the communities which most normalize many rapidly updating packages are going to be at greater risk. That’s not a simple trade off — that enterprise Java app which updates on a decadal cadence is still worse — but it means you need to accept the risk and use other mitigations.
I just don't think that this is that unique to javascript, it's absolutely not about npm, and I don't think that this is well supported as a relevant feature that leads to these attacks.
Ok, I can agree it is a boring comment, but who is worse?
NPM gets targeted both because it is popular and because there is a wider attack surface (lots of little packages promoted by a huge variety of users) I have a high schooler who published work a couple weeks ago. This is good, but it comes with downsides. Maybe a couple more speed bumps or classifiers would be helpful. Maybe a consolidation of under maintained projects and deprecation is in order.
Arguably crates.io is worse. NPM has cooldowns and has for a while, it has had Trusted Publishing for longer, it has human-approved releases that separate CI/CD from actual publishing. Ruby is probably worse in every way.
Yeah, my point is just that other package managers aren't in a great spot. NPM even lets you separate out "publish" and "release" now where you can publish to the registry but you have to separately "ack" that to release. That's kinda a huge win if people use it.
I just think the framing that npm is so bad is really flatly invalid.
I don’t think npm itself is measurably bad in ways that are useful to index on. Like others, I think that JS as an ecosystem has a culture that encourages both dependency sprawl and living at the edge, both of which contribute to the end problem of OSS malware.
(Rust has a similar culture, to be clear. I don’t think it’s a death knell.)
Other than what happened with 'xz', which was upstream of it getting packaged, how many times has this happened in the debian packages system? Also very popular.
Debian's scale for package distribution is tiny and explicitly curated by maintainers. Everything funnels through Debian. The goals are completely different. Debian packages what's necessary for an OS, npm packages everything needed for any projects arbitrarily. Auditing scales to one of those, not the other.
The quantity of packages that exist at 162150 seems to say to me that there's a great many more than what are "necessary" for the operating system to exist.
sudo apt-cache stats
Total package names: 162150 (5189 k)
Total package structures: 145852 (6417 k)
Normal packages: 69505
Pure virtual packages: 1169
Single virtual packages: 64848
Mixed virtual packages: 355
Missing: 9975
Total distinct source versions: 71172 (1708 k)
Total distinct versions: 71172 (6334 k)
Total distinct descriptions: 139719 (3353 k)
Total dependencies: 435057/121405 (10.6 M)
Total ver/file relations: 73378 (1174 k)
Total Desc/File relations: 142692 (2283 k)
Total Provides mappings: 70533 (1693 k)
Total globbed strings: 268430 (6560 k)
Total slack space: 77.3 k
Total space accounted for: 47.0 M
Total buckets in PkgHashTable: 196613
Unused: 93545
Used: 103068
Utilization: 52.4218%
Average entries: 1.4151
Longest: 19
Shortest: 1
Total buckets in GrpHashTable: 196613
Unused: 86025
Used: 110588
Utilization: 56.2465%
Average entries: 1.46625
Longest: 7
Shortest: 1
162k vs millions. That's not even getting into package update velocity, authorship, the totally divergent goals, etc. I just think it's utterly pointless to compare.
No other package manager is worse than NPM. Outside of its 'popularity', there are several fundamental reasons why this continues to happen to NPM:
- Imported packages are not pinned by default.
- Typescript / Javascript's lack of a standard library encourages the developer to import more packages into their codebase to address the short-comings which increases the risk of importing a bad package.
- Post install scripts execute external code by default upon downloading dependencies.
All of this comes by default in the ecosystem and we continue to see more shai-hulud worms all easily targeting NPM. Not even signed packages are enforced by default either.
What do you mean? The lockfile of all package managers is there for pinning the exact versions. For yarn and pnpm, installs on CI run automatically from lockfile only, for npm I think you still need to run `npm ci` instead of `npm install`. But this guarantees that no new versions get pulled automatically in by CI.
> Typescript / Javascript's lack of a standard library
That's true, but it's not an issue of the package manager / registry
> Post install scripts execute external code by default upon downloading dependencies.
They are disabled by default in all package managers now, the user needs to manually allow them
And Russia doesn't give a damn, as they literally gave APT28/29 the mandate to do this, the only exception being that no former Sovjet territories can be attacked.
1. Provenance did not fail, it worked as designed and still shipped malware. The initial 11 packages were published through npm OIDC Trusted Publishing with valid SLSA attestations. The attacker compromised the maintainer's GitHub account and let the projects' own release workflows publish. Provenance proves which commit was built, not that the commit was authorized.
2. There is a booby trap on remediation: the worm installs a watcher that fires an attacker payload when the stolen GitHub token gets revoked. Remove the token monitor first, then rotate.
3. Persistence goes beyond node_modules. It writes .claude/settings.json (SessionStart hook) and .vscode/tasks.json (runOn: folderOpen), each re-executing the dropper. Check repo dotfiles too.
4. No C2 domain to sinkhole: exfil endpoints resolve at runtime from an Ethereum contract. Observed domain is npm-cache.com, but the operator can rotate it and push new code to infected hosts.
If you're auditing: look for setup.mjs, a 727,680 byte Math_Symbol.js (math_init.js in the second wave), and "preinstall": "node setup.mjs" in package.json. Careful, regenerate-unicode-properties ships a legitimate 1 KB Math_Symbol.js; the malicious one is over 700 KB.
Full IOC and package list in the post, updated as the campaign develops.
The cheapest mitigation almost nobody deploys: a version cooldown. These worms get caught fast — this one was flagged same-day, and the article's own timeline shows detection racing ahead of spread. If your CI simply refuses to adopt any version published in the last N days (Renovate supports this natively via minimumReleaseAge), you convert "worm spreads through the ecosystem in hours" into "worm must survive N days of public scrutiny before it can reach you." You give up almost nothing: how often does your product genuinely need a dependency version that's 48 hours old?
Combine that with the workflow split insanitybit describes — build/test jobs holding zero publish credentials, a separate publish job that only touches a finished artifact — and the wormable path is mostly closed without waiting for npm to redesign itself.
None of this is "sufficient" in rcxdude's sense, and that's fine. Sufficiency isn't the bar during an active outbreak; raising the attacker's cost per hop is.
First point.. plus, OTP/MFA authenticated publishing.
This OTP/MFA should come from package repositories, before the package is made publicly available. This is needed so that CD stage is not blocked.
OTP/MFA should be scoped to publishing user/org, not the package. How the OTP/MFA client is managed across the maintainers/org, lies in the scope of maintainers/org.
Updates should be changed to delete + publish, and either should require OTP/MFA. You don't need artificial cooldown if you add a manual, informed action in-between. All these publishes went uninformed to their maintainers.. that's the issue.
I think npm can use chatgpt/claude for each published package to detect these types of attack? And if it sees they can restrict the package from making any changes.
aikido.dev: SAST, AI code analysis and therefore a website that uses 100% CPU and prevents scrolling.
To the point of the article: I don't know why GitHub still allows the release feature. It is complete insanity. Tar archives must be constructed manually and checked for leaked keys etc.
I have a suspicion that a lot of these supply chain compromises are done by the security researchers at security vendors, selling software to protect the software supply chain. Spreading fear to create demand for their products.
Like in the good ole days of Windows 98 and antivirus era, a lot of advanced virus techniques in the wild came from the people who used to work for AV companies
That is not a suspicion, but a (baseless) theory. I mean, your proof is from Windows 98. Might as well been an outlier or urban legend, who knows.
My theory is that people in countries without extradition laws to USA are going to spray the shit out of Americans. With spam, and crap like this worm, troll farms, and the like.
Also baseless, but people who work at SOC can notice an increase since some changes happened in geopolitics this decade.
> Update — August 4, 2026, 13:37 CEST: At least 868 packages (across 1381 versions) have been compromised by the worm, with a combined total of over 2 billion monthly installs at the time of writing.
This is such lazy or click baiting writing. Who cares how many installations there are per month normally? The high install numbers are almost certainly from running in CI where such secrets don’t exist. How many installs actually occur in a non CI environment and of those how many were the compromised version?
CIs having per-package installs : version ratio > 1 are fundamentally dumb idea in the first place.
In a way, this teaches us that there is something as too much reliability in an ecosystem. Specifically, Github is apparently reliable enough that its occasional outages are insufficient to kill all the companies running these dumb CI setups that redownload packages from the Internet every build - so there's no incentive to reduce waste and improve security globally.
Redownloading and automatically updating to the new version are pretty orthogonal. You can have a solution which caches the packages but would still update to a new version or a solution which redownloads the (verifiably) same package each time. Vendoring libraries stops both but it's the automatic updates which are the biggest risk factor.
It's time pre-install / post-install hooks were killed off. Start with a moratorium on any new ones.
reply