Sadly JS has ways around it that is far from obvious since you can chain effects over multiple files that leads to running code.
Like the following example (you can paste it into node to verify), could be spread out over multiple source files to make it even harder to follow:
// prelude 1, obfuscate the constructor property name to avoid raising simple analyser alarms
const prefix = "construction".substring(0,7);
const suffix = "tractor".substring(3);
const obfuscatedConstructorName = prefix + suffix; // innocent looking, but we have the indexing name.
// prelude 2, get the Function class by indexing a function object with our constructor property name (that does not show up in source-code)
const existingFunction = ()=>"nothing here";
const InnocentLookingClass = existingFunction[obfuscatedConstructorName];
// payload decoding elsewhere (this is where we decode our nasty source)
const nastyPayloadDisguisedAsData = "console.log('sourced string that could be malicious')";
// Unrelated location where payload gets executed
const hardToMissFun = new InnocentLookingClass(nastyPayloadDisguisedAsData);
hardToMissFun(); // when this function is run somewhere.. the nasty things happen.
Unless you have a data-tracing verifier or a sandbox that is continiously run it's going to be very hard to even come close to determining that arbitrary code is being evaluated in this example. Not a single trace of eval or even that the property name constructor is used.
IMO while the bar is high to say "it's the responsibility of the repository operator itself to guard against a certain class of attack" - I think this qualifies. The same way GitHub provides Secret Scanning [0], it should alert upon spans of zero-width characters that are not used in a linguistically standard way (don't need an LLM for this, just n-tuples).
Sure, third-party services like the OP can provide bots that can scan. But if you create an ecosystem in which PRs can be submitted by threat actors, part of your commitment to the community should be to provide visibility into attacks that cannot be seen by the naked eye, and make that protection the norm rather than the exception.
I think a "force visible ASCII for files whose names match a specific pattern" mode would be a simple thing to help. (You might be able to use the "encoding" command in the .gitattributes file for this, although I don't know if this would cause errors or warnings to be reported, and it might depend on the implementation.)
It baffles me that any maintainer would merge code like the one highlighted in the issue, without knowing what it does. That’s regardless of being or not being able to see the “invisible” characters. There’s a transforming function here and an eval() call.
The mere fact that a software maintainer would merge code without knowing what it does says more about the terrible state of software.
In this instance the PR that was merged was from 6 years ago and was clear https://github.com/pedronauck/reworm/pull/28. Looks to me like a force push overwrote the commit that now exists in history since it was done 6y later.
> It baffles me that any maintainer would merge code like the one highlighted in the issue, without knowing what it does.
I don't know if it is relevant in any specific case that is being discussed here, but if the exploit route is via gaining access to the accounts of previously trusted submitters (or otherwise being able to impersonate them) it could be a case of teams with a pile of PRs to review (many of which are the sloppy unverified LLM output that is causing a problem for some popular projects) lets through an update from a trusted source that has been compromised.
It could correctly be argued that this is a problem caused by laziness and corner cutting, but it is still understandable because projects that are essentially run by a volunteer workforce have limited time resources available.
Looks like the repo owner force-pushed a bad commit to replace an existing one. But then, why not forge it to maintain the existing timestamp + author, e.g. via `git commit --amend -C df8c18`?
Unicode should be for visible characters. Invisible characters are an abomination. So are ways to hide text by using Unicode so-called "characters" to cause the cursor to go backwards.
Things that vanish on a printout should not be in Unicode.
That ship has sailed, but I consider Unicode a good thing, yet I consider it problematic to support Unicode in every domain.
I should be able to use Ü as a cursed smiley in text, and many more writing systems supported by Unicode support even more funny things. That's a good thing.
On the other hand, if technical and display file names (to GUI users) were separate, my need for crazy characters in file names, code bases and such are very limited. Lower ASCII for actual file names consumed by technical people is sufficient to me.
No need to remove them. Just make them visible for applications that don't need to render every language. Make that behavior optional as well in case you really want to name characters with Hangul or Tibetan.
Some middle ground so that you can use greek letters in Julia might be nice as well.
But I don't see any purpose in using the Personal Use Areas (PUA) in programming.
Attacks employing invisible characters are not a new thing. Prior efforts here include terminal escape sequences, possibly hidden with CSS that if blindly copied and pasted would execute who knows what if the particular terminal allowed escape sequences to do too much (a common feature of featuritis) or the terminal had errors in its invisible character parsing code.
For data or code hiding the Acme::Bleach Perl module is an old example though by no means the oldest example of such. This is largely irrelevant given how relevant not learning from history is for most.
Invisible characters may also cause hard to debug issues, such as lpr(1) not working for a user, who turned out to have a control character hiding in their .cshrc. Such things as hex viewers and OCD levels of attention to detail are suggested.
I wonder if this could be used for prompt injection, if you copy and paste the seemingly empty string into an LLM does it understand? Maybe the affect Unicode characters aren’t tokenized.
I use non-Unicode mode in the terminal emulator (and text editors, etc), I use a non-Unicode locale, and will always use ASCII for most kind of source code files (mainly C) (in some cases, other character sets will be used such as PC character set, but usually it will be ASCII). Doing this will mitigate many of this when maintaining your own software. I am apparently not the only one; I have seen others suggest similar things. (If you need non-ASCII text (e.g. for documentation) you might store them in separate files instead. If you only need a small number of them in a few string literals, then you might use the \x escapes; add comments if necessary to explain it.)
The article is about in JavaScript, although it can apply to other programming languages as well. However, even in JavaScript, you can use \u escapes in place of the non-ASCII characters. (One of my ideas in a programming language design intended to be better instead of C, is that it forces visible ASCII (and a few control characters, with some restrictions on their use), unless you specify by a directive or switch that you want to allow non-ASCII bytes.)
> ... and will always use ASCII for most kind of source code files
Same. And I enforce it. I've got scripts and hooks that enforces source files to only ever be a subset of ASCII (not even all ASCII codes have their place in source code).
Unicode chars strings are perfectly fine in resource files. You can build perfectly i18n/l10n apps and webapps without ever using a single Unicode character in a source file. And if you really do need one, there's indeed ASCII escaping available in many languages.
Some shall complan that their name as "Author: ..." in comments cannot be written properly in ASCII. If I wanted to be facetious I'd say that soon we'll see:
Why can't code editors have a default-on feature where they show any invisible character (other than newlines)? I seem to remember Sublime doing this at least in some cases... the characters were rendered as a lozenge shape with the hex value of the character.
Is there ever a circumstance where the invisible characters are both legitimate and you as a software developer wouldn't want to see them in the source code?
I feel like the threat of this type of thing is really overstated.
Sure the payload is invisible (although tbh im surprised it is. PUA characters usually show up as boxes with hexcodes for me), but the part where you put an "empty" string through eval isn't.
If you are not reviewing your code enough to notice something as non sensical as eval() an empty string, would you really notice the non obfuscated payload either?
Honestly I was expecting more. There are many languages that support Unicode in variable or function names and I expected it to be used there.
It sounds like Python only allows approved Unicode characters to start a variable name but if it allowed any you could do something like `nonprintable = lambda x: insert exploit code here`. If that was hidden in what looked like a blank line between other additions would you catch it?
I'm sure there's some other language out there that has similar syntax and lax Unicode rules this could be used in.
The solution is that this and many other Unicode formatting characters should be ignored and converted to a visible indicator in all code views when you expect plain text.
Invisible characters, lookalike characters, reversing text order attacks [1].. the only way to use unicode safely seems to be by whitelisting a small subset of it.
And please, everyone arguing the code snippet should never have passed review - do you honestly believe this is the only kind of attack that can exploit invisible characters?
My hot take is that all programming languages should go back to only accepting source code saved in 7-bit ASCII. With perhaps an exception for comments.
The exact quote is "Thanks for the submission! We have reviewed your report and validated your findings. After internally assessing your report based on factors including the complexity of successfully exploiting the vulnerability, the potential data and information exposure, as well as the systems and users that would be impacted, we have determined that they do not present a significant security risk to be eligible under our rewards structure." The funny thing is, they actually gave me $500 and a lifetime GitHub Pro for the submission.
Tangential, but that's quite interesting, I had no idea you could get GitHub Pro for life, and certainly not through something as "accessible" as bug bounties.
57 comments
[ 2.8 ms ] story [ 67.7 ms ] threadLike the following example (you can paste it into node to verify), could be spread out over multiple source files to make it even harder to follow:
Unless you have a data-tracing verifier or a sandbox that is continiously run it's going to be very hard to even come close to determining that arbitrary code is being evaluated in this example. Not a single trace of eval or even that the property name constructor is used.Sure, third-party services like the OP can provide bots that can scan. But if you create an ecosystem in which PRs can be submitted by threat actors, part of your commitment to the community should be to provide visibility into attacks that cannot be seen by the naked eye, and make that protection the norm rather than the exception.
[0] https://docs.github.com/en/get-started/learning-about-github...
i see squares on a properly configured vim on xterm.
https://docs.github.com/en/authentication/managing-commit-si...
I first heard about the possibility of this kind of attack >10 years ago, and I'll sometimes do a xxd if i'm feeling a bit paranoid.
See commenter on their 2025 bounty for reporting it, won't-fix resolution: https://news.ycombinator.com/item?id=47393393
I am wondering how that they've LLM, are people using them for making new kind of malicious codes more sophisticated than before?
The mere fact that a software maintainer would merge code without knowing what it does says more about the terrible state of software.
I don't know if it is relevant in any specific case that is being discussed here, but if the exploit route is via gaining access to the accounts of previously trusted submitters (or otherwise being able to impersonate them) it could be a case of teams with a pile of PRs to review (many of which are the sloppy unverified LLM output that is causing a problem for some popular projects) lets through an update from a trusted source that has been compromised.
It could correctly be argued that this is a problem caused by laziness and corner cutting, but it is still understandable because projects that are essentially run by a volunteer workforce have limited time resources available.
Innocuous PR (but do note the line about "pedronauck pushed a commit that referenced this pull request last week"): https://github.com/pedronauck/reworm/pull/28
Original commit: https://github.com/pedronauck/reworm/commit/df8c18
Amended commit: https://github.com/pedronauck/reworm/commit/d50cd8
Either way, pretty clear sign that the owner's creds (and possibly an entire machine) are compromised.
Things that vanish on a printout should not be in Unicode.
Remove them from Unicode.
I should be able to use Ü as a cursed smiley in text, and many more writing systems supported by Unicode support even more funny things. That's a good thing.
On the other hand, if technical and display file names (to GUI users) were separate, my need for crazy characters in file names, code bases and such are very limited. Lower ASCII for actual file names consumed by technical people is sufficient to me.
Also this attack doesnt seem to use invisible characters just characters that dont have an assigned meaning.
Some middle ground so that you can use greek letters in Julia might be nice as well.
But I don't see any purpose in using the Personal Use Areas (PUA) in programming.
For data or code hiding the Acme::Bleach Perl module is an old example though by no means the oldest example of such. This is largely irrelevant given how relevant not learning from history is for most.
Invisible characters may also cause hard to debug issues, such as lpr(1) not working for a user, who turned out to have a control character hiding in their .cshrc. Such things as hex viewers and OCD levels of attention to detail are suggested.
My clawbot & other AI agents already have this figured out.
/s
The article is about in JavaScript, although it can apply to other programming languages as well. However, even in JavaScript, you can use \u escapes in place of the non-ASCII characters. (One of my ideas in a programming language design intended to be better instead of C, is that it forces visible ASCII (and a few control characters, with some restrictions on their use), unless you specify by a directive or switch that you want to allow non-ASCII bytes.)
Same. And I enforce it. I've got scripts and hooks that enforces source files to only ever be a subset of ASCII (not even all ASCII codes have their place in source code).
Unicode chars strings are perfectly fine in resource files. You can build perfectly i18n/l10n apps and webapps without ever using a single Unicode character in a source file. And if you really do need one, there's indeed ASCII escaping available in many languages.
Some shall complan that their name as "Author: ..." in comments cannot be written properly in ASCII. If I wanted to be facetious I'd say that soon we'll see:
and so the point shall be moot anyway.The biggest use of Unicode in source repos now might be LLM slop, so I certainly don't miss its absence at all.
Are people using eval() in production code?
Is there ever a circumstance where the invisible characters are both legitimate and you as a software developer wouldn't want to see them in the source code?
Sure the payload is invisible (although tbh im surprised it is. PUA characters usually show up as boxes with hexcodes for me), but the part where you put an "empty" string through eval isn't.
If you are not reviewing your code enough to notice something as non sensical as eval() an empty string, would you really notice the non obfuscated payload either?
It sounds like Python only allows approved Unicode characters to start a variable name but if it allowed any you could do something like `nonprintable = lambda x: insert exploit code here`. If that was hidden in what looked like a blank line between other additions would you catch it?
I'm sure there's some other language out there that has similar syntax and lax Unicode rules this could be used in.
The solution is that this and many other Unicode formatting characters should be ignored and converted to a visible indicator in all code views when you expect plain text.
This isn't about formating characters, this is about private use characters.
And please, everyone arguing the code snippet should never have passed review - do you honestly believe this is the only kind of attack that can exploit invisible characters?
[1] https://attack.mitre.org/techniques/T1036/002/
Though also seems relatively straightforward to code up in your scripting language of choice to call from pre-commit.
Of course, it doesn't work though. I reported this to their bug bounty, they paid me a bounty, and told me "we won't be fixing it": https://joshua.hu/2025-bug-bounty-stories-fail#githubs-utf-f...
The exact quote is "Thanks for the submission! We have reviewed your report and validated your findings. After internally assessing your report based on factors including the complexity of successfully exploiting the vulnerability, the potential data and information exposure, as well as the systems and users that would be impacted, we have determined that they do not present a significant security risk to be eligible under our rewards structure." The funny thing is, they actually gave me $500 and a lifetime GitHub Pro for the submission.