We found a division by zero bug in FFmpeg with a vibecoded fuzzer
code.ffmpeg.org183 points by dclavijo 8 hours ago
183 points by dclavijo 8 hours ago
A patch for this was submitted in April: https://lists.ffmpeg.org/archives/list/ffmpeg-devel@ffmpeg.o...
Edit: And there was discussion about this back in 2024 as well
Oddly enough I can’t access that site, it just heats up my phone solving hashes. Gave up after about a minute and anubis had only made it less than halfway through.
I doubt the real bots have any trouble bypassing it.
It's puzzling how mild the reactions are to Anubis compared to the people reacting to seeing one singular Cloudflare captcha checkbox. I'd much rather a checkbox than a brief CPU-intensive hashing session.
I think when people are complaining about Captcha they're complaining about yet another "pick 6-20 pictures of traffic lights/school busses/stairs/stop signs/bicycles."
If they want to train an AI they should pay for it like everyone else. Modern bots have zero problems solving these, it's just free training for them.
It also took insanely long on my iPhone 16. "Made with heart in Canada" but configured poorly.
Yeah, is it trying to mine bitcoin or something? Anubis usually takes a second but here I waited a minute and got 20% through on a modern phone.
Presumably they've configured it to use a higher difficulty challenge due to high rates of scraping on their bugtracker
Difficulty 6 which some parts of FFmpeg use, is about the highest difficulty you can assign with the default Anubis config. For me personally I only serve that difficulty if I'm near certain the user is a bot. Serving it to everyone sure is a choice.
happy to report my bots have no trouble with anubis or any other pow mechanism, little bit of deno and i'm right through.
Took less than a minute in my 5 year old xiaomi phone. It did take way longer than other Anubis sites I've seen.
Yeah, it's painful.
I get this crap when browsing on desktop a lot as well, principally because I stubbornly use Firefox as my main browser, and I habitually use a VPN when I connect my laptop to unsecured or even secured-but-accessible-to-large-numbers-of-people WiFi networks.
Like, seriously, bot detection "specialists", fuck off: I'm not a bot but your bot detection software IS shit, and I DO resent your shit software draining my battery and getting in my way. Learn to do your jobs properly, will you?
And don't come crying to me about how the problem you're trying to solve is "hard". I don't care: you chose it, you chose to considerably worsen the web browsing experience of millions of people globally, nobody made you. So go and find a different job if you're incapable of doing the one you have.
And if it's so "hard" why does your entire solution seem to be predicated on anyone's a bot if they're not running Chrome, or they are running an adblocker, or they appear to be from an unusual country that doesn't match their system language? Seriously, is this the level of sophistication you hacks operate at? To solve your "hard" problem?
You are extremely lame. Get out of my way.
> And don't come crying to me about how the problem you're trying to solve is "hard". I don't care: you chose it, you chose to considerably worsen the web browsing experience of millions of people globally, nobody made you.
Unfortunately, if you let all the bots in, they overwhelm your servers, and then nobody can access the website.
Not if you use a decentralized peer-to-peer Git forge like https://radicle.network. If one node goes down, users can still access the same issues/PRs from another endpoint.
I assume you're offering to pay for the increased server costs?
I had some git hosting up for a while, and was serving hundreds of qps and several terabytes per month. I can only imagine want significant sites are serving.
For those with difficulty accessing:
- - -
From: Anthony Hurtado <[redacted since hn has no scrape protection]>
vpk_read_packet() divides vpk->last_block_size and (par->block_align - vpk->last_block_size) by par->ch_layout.nb_channels without checking for zero.
While vpk_read_header() validates nb_channels > 0, the codec parameters may become zero through format probing misidentification (VPK probe score is 2/3 of AVPROBE_SCORE_MAX) or codec parameter reset, causing SIGFPE.
Fix by:
- Checking nb_channels != 0 before division in vpk_read_packet
- Returning EOF for empty last blocks (last_block_size == 0)
- Validating block_count > 0 in vpk_read_header
- Validating last_block_size <= block_align in vpk_read_header
Found by fuzzing with libFuzzer + AddressSanitizer. Reproduces with 10 distinct inputs.
[patch redacted for brevity]
Thank you! I gave up after more than 2 whole minutes of waiting on a high-end smartphone. I'm not sure this keeps bots out, but it definitely keeps users out…
A patch was submitted, but apparently not merged. That was also my experience trying to submit a patch for https://trac.ffmpeg.org/ticket/8738 . Somebody on the bug tracker took note, but was apparently unable to effect a merge in the intervening years.
Maybe now that ffmpeg is using Forgejo, the ball won't be dropped like this as often. Or there'll just be a five-digit number of open pull requests instead.
It’s interesting how AI may both raise and lower the quality of software. It’s very easy to send an AI agent on an open-ended bug hunt, and if it wastes a bunch of time and effort and finds nothing, no big deal. Time is much more important for a human developer with a salary.
Finding the bugs with LLMs is easy. Reviewing the output, cleaning it up, and making sure it doesn't break something else is the hard part.
This is where I believe strong typing (like, Haskell-strong or stronger) and functional programming in general will be a win. The confidence I have that my fixes are localised when fixing Haskell code is infinitely stronger than fixing even Java, not speak about C, code.
Haskell's type system would not easily prevent this bug. It's not good at numeric/logic issues like that. When people say "Haskell makes it impossible to write bugs" they mean "Haskell has enums" (ADTs).
Liquid Haskell might require you to prove that the divisor is nonzero, but even in standard Haskell there's common idioms for ensuring that a list is non-empty (data NonEmpty a = a :| [a]) or that text is non-empty (newtype NonEmptyText = NonEmptyText Text, with non-exported constructor, helpers like make :: Text -> NonEmptyText, or more advanced tricks like https://exploring-better-ways.bellroy.com/haskell-koan-type-... ).
The big problem preventing this approach from working for numbers is that it's just so cumbersome there. Most of this is because all the arithmetic operators are bundled into a single Num typeclass, and `fromInteger :: Num a => Integer -> a` has a type that's impossible for a "non-zero number" wrapper to satisfy.
Definitely room for improvement on Haskell's standard library when it comes to the number-related type classes. Modern Haskell could do very well in this area with a good type-class redesign in this area. The issue I think is that this would invalidate a lot of existing code, relying upon that. But you can already replace Prelude with something else in your own code if you want to.
OOP has those too, and they're very annoying.
In Haskell they are a little less annoying. It is just easier to reason about (including proving) pure functions.
I meant the constrained types by hiding the constructors. Super annoying, not automatically convertible, in Haskell you have to remember what the fake constructor is called, and write it every time you use it, but at least it's efficiently implemented with newtype, unlike the Java OOP version. Think about writing a value with several nested constrained types, like NonEmptyListOne (makeNonZeroNumber 42, 'h' `NonEmptyString` "ello world"). It's just really annoying.
The blog link I mentioned avoids this cost with literals, by providing using a required type argument to check the string length at compile time without TH. It requires a relatively recent GHC:
make :: forall symbol -> (IsNonEmptySymbol symbol) => NonEmptyText
type family IsNonEmptySymbol symbol :: Constraint where
IsNonEmptySymbol "" = Unsatisfiable (Text "Expected a non-empty string")
IsNonEmptySymbol _ = (()::Constraint) -- empty constraint is always satisfiedI am not claiming you cant write buggy code in Haskell! But following good functional style, your bug will more likely be compartmentalised, and fixing it will not break some other part of your program.
You can write good functional code in many languages. (Even C++!)
Sure! I have done my fair share of pretending Java and C++ support my functional style. But at the end of the day, you have better support for writing that style in a real functional programming language. And I wonder how well one can enforce a functional style in say Java or C++ upon the LLMs. Who knows, they might be great at it?
Imo, formal methods like more expressive/stricter type systems are key to making LLM generated code successful. Of course models will get better, but trusting the output will become much easier with a type system that proves more properties.
What's stronger than Haskell?
Dependent types is one possible direction. Not sure when a language with dependent types will arise which will be useful for making real programs.
Agda is the most mature dependently typed programming languae (having been around since the 90s – it is basically Haskell on steroids), but has a more proof-assistant flavor than an actual programming language flavor. Opus & Fable write Agda quite well, so LLMs can understand dependent types.
Anything with ranged numeric types. Like everyone's favorite functional programming language, Ada.
This issue raises SIGFPE. Ada would raise Constraint_error, which is easier to catch than a signal, but still occurs at runtime.
You need range proofs to be 100% safe, and then you can as well use the regular type because invalid values will not occur.
If finding the bugs with LLMs is easy. Then making sure it doesn't break something else is just LLMs finding no bugs. Easy.
That hasn’t been that bad. My real issue has been the time sink involved in following along with the maintainer and jumper through their hoops. Even after I demonstrate a flaw and a potential fix. My schedule is just so busy I need to pencil in time to deal with them.
The missing part of this is that verifying the bug with LLMs is also easy, and so is adversarially reviewing the proposed fix with LLMs.
The only thing left for you to do should be directional decisions. The LLMs should pause and rope you in if the fix involves directional/invariant changes.
No one can keep up with the volume of code AI produces.
We wont stop using AI.
We will use AI to check AI.
Of course this is crazy, but it will also unlock pretty insane scaling and productivity and ultimately we will manage it on either end via requirements and tests.
> it will also unlock pretty insane scaling and productivity
Insane scaling of bloat, bugs, and technical debt I'd say.
> We will manage it on either end via requirements and tests
It is so crazy that this is being touted as a sane strategy. When I was a much worse programmer, I tried to write a big complicated string manipulation function to take two types of scripts in a language and add diacritics. I had the requirements very clear. I had the tests very clearly with all the edge cases. But I didn't have a good and clear picture of how to attack the problem which was quite novel for me. As I got closer to passing all the tests it got exponentially more unruly and confusing. And nearing the end I was frantically changing little bits here and there wincing and praying and hoping the tests would pass. "Please work! Come on!" Then when I got close enough, I could never ever think about touching that mess again.
I was a below average programmer then throwing myself at some novel problem I didn't understand. Throwing LLMs that produce below average code at novel problems and relying on tests and requirements is not where we want to go to make real progress.
(Years later after much learning and coding myself I was able to redo the function in a totally different way. This time I actually understood how to attack the strange problem and made something clean, clear, and robust that just worked. The tests then become a secondary guardrail, not the main force of correction.)
We are seeing such a massive regression from what we've learned over the years of CS.
>Insane scaling of bloat, bugs, and technical debt I'd say.
You just described every legacy codebase. Many of which are widely used and do a lot of sales. You dont need a clean codebase to have a valuable product.
>It is so crazy that this is being touted as a sane strategy.
Re-read what I said. I literally called it crazy.
It is the same dynamic that gave us customer service from some call center in India. Why would companies do this? Customer service got worse. Are they stupid? No, it's just worth it. The quality goes down but the business can scale more so it doesnt matter.
AI will absolutely be good enough at doing things that we'll happily accept some jankiness at times so that we can devote an extra 3000 hours per year per person to other things.
Im not even suggesting its a good thing. I just think the incentive structure dictates it. You're not going to have time to maintain a small slice of some service by hand.
I think all code is technical debt in a way. Good code is a necessary evil, bad code is more evil than necessary.
Generating code automatically when you're not even quite sure what it is or even should be doing is insanity.
I'm not so sure LLM code today is below average. There was a time that things posted to dailywtf were normal everyday stuff
You shared a story of a novice incompetent human programmer and this should tell us that AI is bad at coding.
It's mostly (not entirely, but mostly) finding security issues in old human-written code. It'll eventually start running out of those.
From that standpoint, it's not a crazy setup security-wise. Maybe still crazy for development.
You can point AI at any AI produced code and ask it to review it, get back 10 bullet points and a few pages of prose. And the fun part is, you can do that over and over and over again!
This happens all the time. Yesterday, I ran into an especially egregious case.
I had Fable add a new subcommand to our internal CLI tool. I reviewed and tested it locally and had to suggest several fixes that I feel like I wouldn't have had to tell a human senior engineer to do. When it finally submitted the PR, I had it on a loop waiting a few minutes for comments on the PR, then assessing/addressing/replying-to/resolving them, and then repeating again until all AI reviewers were okay with it. It ended up going through dozens of revisions and ended up with 160 comments left on the PR.
Volume..... <sigh>
It used to be considered a quality of good code that there would be less code, not more.
Some people always tryin to get the highscore on golf.
You're suggesting that LLMs get better at fixing bugs/vulnerabilities, but at the same time stop getting better at finding them? What if this difference is inherent and essential?
> You're suggesting that LLMs get better at fixing bugs/vulnerabilities, but at the same time stop getting better at finding them?
Are you implying that all code writing by LLMs atm is bug-free?
Absolutely not. By most accounts they're terrible at fixing anything other than trivial bugs in complex codebases e.g. Linux kernel, but they're much better at finding them.
In fairness at root this has been going on for awhile. No one can keep up with the volume of machine code that modern more abstracted codebases produce.
We didn't stop using syntactic programming languages we used code to check code.
Not sure it's really crazy at all. It's been an abstraction for programmers probably since we stopped soldering transistors to each other.
There is a MAJOR difference between predictable generated machine code and Russian Roulette code generator.
[flagged]
> LLMs do find bugs, do save time
They find bugs but whether they save time is nowhere near as clear as you try to insinuate here.
In my experience, there are two ways to use AI: speed or quality. Speed is where you give the AI a task to do and you review it; quality is where you write the code yourself and you get AI to review it. Both are valid for different situations.
My plan for bigger things is mostly:
Generate multiple solutions- they do not to work 100% correctly. And than I check which I would prefer. Which is more to our applications taste.
And than I would take the vibe output as a kind of a ‚plan‘ which I use to implement but not follow 100% and at the end I take my solution and review it. I gain speed with that because I often can quickly see the pros and cons of a solution way better than when I would manually do it and hang on a major roadblock and also I even see such roadblocks in the vibe output - it’s mostly the part with an unnecessary amount of new code that looks nonsensical.
Using a LLM whose output is slowed to the rate of a human programmer as a pair programming partner is a very interesting experience.
> It’s very easy to send an AI agent on an open-ended bug hunt, and if it wastes a bunch of time and effort and finds nothing, no big deal.
No big deal? It’s not like it’s free… tokens cost money.
When talking about LLM tokens the cost is almost always being implicitly compared to very expensive human developer time.
I don't care if you call it an over-engineered looping machine or what, there are concrete benefits to using LLMs for this. They work faster than developing your own looping algorithm and more often produce useful results than not.
It's not even like fuzzers are valuable because of the process they use specifically either; the value is that they produce a concrete input that you can use as a reproducible test case at that point. The value could be produced by gazing into a crystal ball for all I care, as long as I can use what it gives me to reproduce a bug.
I dislike AI, but if AI finds real bugs then this is in my opinion objectively a positive thing. Of course the question is what constitutes a real bug.
Unfiltered models will help build exploits for the bugs they find, so there is some means of measuring their efficacy.
If you're just talking about security bugs.
There are also non security bugs that don't have exploits but just make the user experience worse.
A.I. is useful for this. But it would be even more useful if all new code were written in Rust or some other memory-safe language.
A.I. could also be used to port C/C++ codebases to Rust, which isn't economically feasible at the moment.
I mean I get the sentiment but Rust won't save you against division by zero, it'll just panic at runtime like every other language.
From a security perspective, panic at runtime is not that bad for security. Much better than continuing to run with undefined behavior. If someone sends a malformed video in and it crashes the ffmpeg process you can just log it and restart it. Vs potentially exploiting the system.
Whatever about the specifics of this bug and whether its a useful vector, this is not surprising even in the slightest?
My current opinion on LLMs is that they are superhuman in that they lack fatigue, they have close to full knowledge across all subjects which are known to humans at least publicly, and the fact that you can vibe code a harness to look for bugs in a famously complicated C codebase is intern level stuff and hardly news.
Smart aspiring blackhats will be targeting tmux next, both with light llm jailbreaks, light supply chain attacks (web search results) and LPEs within certain environments which weren't particularly useful before but with agents running on auto mode for hours become a very valuable springboard. I'm not sure on the quality of tmux code but I know its written in C and is very complex and was not at all designed to defend against this type of threat.
Not that it doesn’t have issues, but I’m not sure why you’d choose tmux of all things. It runs as a user and has no privileges to escalate. It was written for and is part of OpenBSD and follows their security hardening practices.
(There actually was one privilege escalation bug in tmux, but it actually seems like a distro packaging error. The distro setgid the executable so the resulting shell inherited the additional group. This didn’t require any exploit, that’s just how child process inheritance works.)
as I mentioned in another sibling, its because it's a very common denominator in high value targets. I didn't know its legacy was from OpenBSD but I really doubt that that helps it much in this scenario, when I say LPE I'm not talking about user to root elevation, I'm talking parsed text/control sequences to arb code execution in the user context. These will slip past llm classifiers as safe and I'm fairly sure that they are extremely common in codebases like tmux, despite them having strong security posture its just a threat that was previously a bit outlandish and not accounted for.
persisted malicious code running in your tmux process that you don't know about is probably not where you want to be, for obvious reasons.
Agent harnesses aren’t going to output ansi escape sequences verbatim to the terminal.
If you wanted to booby trap a repository then you’re far better off with a prompt injection attack.
I don’t think tmux is the most worthwhile target because you’d need the user to either execute code locally (thus negating any point in targeting tmux) or rely on the user curl or cat some compromised document (in which case you’re better off targeting curl or cat).
the point is tmux is being used by many developers working in high value targets to automate long running unsupervised agent tasks. you don't need the user to execute code, you need _their agent_ to stumble on the wrong search result or github repo and it wont be noticed for hours that they loaded a persistent threat into your environment.
That seems even harder to do because an agent wouldnt be output text verbatim, which means you cant make use of a rendering bug (eg parsing escape codes).
So you’re back to depending on the agent to execute code locally. at which point you’ve already compromised the system so don’t need a tmux bug.
I’ve spent a lot of time in tmux. Including writing a frontend for it. So I’m probably more familiar than most. And I hear a lot of people say tmux (specifically) is a vulnerability because it’s written in C. But I struggle to see how it’s any more of a vulnerability than (for example) coreutils. Or any other piece of software for that matter.
the fact that you can vibe code a harness to look for bugs in a famously complicated C codebase is intern level stuff and hardly news
It seems like this would have been pure fantasy not that long ago though. So why isn’t it noteworthy again? I don’t really follow what you’re complaining about.
This is not a real bug in FFmpeg. This is a demonstration that if you control a custom AVIO module it is possible to crash FFmpeg by giving it bad data.
Not custom. It's an existing module for a format called VPK. It's a quite trivial bug though, not exploitable apart from DOS and won't ever happen in a real file.
I even question if it is a DOS vector. So the thread crashes and then the system that controls the threads cleans it up and opens a new thread. Seems to be a trivial impact, unless it locks up the thread somehow.
Threads don't work that way. A fatal exception on any thread kills the process.
And the parent will spawn a new process. Unless the server is terribly poorly misconfigured.
No doubt fuzzers (vibecoded or otherwise) can be powerful, but can't you just mark all "/" as potential divide by zero errors?
I guess sometimes developers think they "know" some variable won't be zero, but unless it checked explicitly or by the compiler, that shouldn't be trusted.
> but can't you just mark all "/" as potential divide by zero errors?
If you’re accepting large false positives rates: yes.
If you want users to take your warnings serious: no.
(Nitpick: you certainly don’t want to flag _all_ of them. Divisions by non-zero constants definitely should be excluded, for example (integer division by -1 can lead to overflow, but that would be a different warning))
What are you suggesting and how would it be different than how SIGFPE already works?
If it's possible for program execution with some particular input to lead to a divide-by-zero, that's a bug, especially if the program is expected to be able to handle malformed inputs, or perhaps even deliberately malicious ones. It's not trivial to determine whether a program does this correctly. If it was, program analysis would be easy.
Division can 'go wrong' for certain inputs, but it's not just division. In C, signed integer addition, subtraction, and multiplication, all give undefined behaviour on overflow.
As 'Someone' already pointed out, it's not helpful to just flag all uses of the division operator, or of other potentially dangerous operators. Minimising false positives is one of the core challenges of program analysis.
I mean there could be a guard clause? But yeah, seems like this could be statically evaluated like how some IDEs see a null check and don’t complain about nullability within the same scope.
Am I missing something? Who cares? This isn't a security issue, it's just an unexploitable crash on bad data.
No, you're not. It's a minor bug, probably with an easy fix, that deserves to be fixed. It's not worthy of front page HN...
Why submit an issue rather than just making the fix and adding the tests in PR? Seems like they're just making work for the maintainers.
OP here: A bug report just needs a proof of existence for the condition while a bug fix needs a proof of correctness. Sometimes is the best to let the developers who are day to day in the codebase to choose the best fix and if they what to fix it.
Funny thing, I know I'm brushing up against something in gStreamer developer, but Fable flips out. I have only a loose idea where the issue might be lurking.
Next week, I'll apply for the cyber and I suspect I'll find something similar.
Right now, it's just annoying and thanks the OpenAI cyber was much easier to get access to.
It is interesting that FFmpeg has its own Git server. Maybe we should move there too?
Lots of projects run their own git or forgejo or similar. I run my own private forge, and it has a higher uptime than GitHub. (A shockingly low bar, tbh)
It’s surprisingly simple to setup, and the hardware requirements are pretty small for a private or small forge, as it’s usually a relatively small number of users/repos/etc.
> It is interesting that FFmpeg has its own Git server. Maybe we should move there too?
Git is a DVCS. I know many people only ever used Git through Github and forgot what the 'D' in DVCS means but whether or not they remember what the 'D' stands for, running your own Git server is trivial. Especially in this day and age of LLMs were you can just ask: "Clone this repo and convert it to base Git repo and serve it on the LAN PLZ KTHX".
The result is going to be more stable than Github and, arguably, more secure too.
If you have SSH access to a server and Git is installed on that server, you can use it as a Git server. No additional setup is required. The Git client knows how to log in and invoke the Git server over SSH.
What we need is a numeric type that cannot be zero.
What we need are refinement types, where there’s a base type and a predicate. F* has this:
val (/) : int -> (divisor:int { divisor <> 0 }) -> intAnd also cannot be INT_MIN, otherwise -1 / INT_MIN is undefined behaviour(!) in C and C++.
The only way to achieve this is to either put a runtime software check on a variable whenever it's assigned/used, or to literally add hardware support in processors themselves which literally throws an interrupt when a "neverShallBeZero" variable is assigned to zero.
There's no viable way to statically prove at compile-time that these variables will never become zero at runtime, ultimately forcing a system of endless runtime checks (be it software or hardware)... which is why processors already throw exception interrupts when division by zero is attempted.
You're kind of saying the only way to do it is in software or hardware :)
An alternative https://en.wikipedia.org/wiki/Projectively_extended_real_lin...
The projectively extended real line defines division by zero, no reason you couldn't have a floating point type that implemented it.
>There's no viable way to statically prove at compile-time that these variables will never become zero at runtime
strongly typed programming languages like Ada allow for types which have ranges such as disallowing zero -- but also any arbitrary thing like you can create a floating point "degrees" type which is [0.0, 360.0] or any other ranged type
It would be more flexible for a compiler to reuse the range analysis logic used in optimizations for statically verifiable divide by zeros. That way you could extend it to other things like statically verifiable overflows.
For stuff like niche value optimization sure. For practical arithmetic code, nah. Like with this bug, all that changed is that garbage data in gives the user an error that they tried to process garbage data. Adding a new type doesn't make the code better, it just moves the error around. And you really don't want an infix division operator to fail to type check if the right hand side isn't a nonzero type, do you?
I find it pretty cool that a fuzzer thrown together this way actually found a bug in ffmpeg
note that this seems to be a bug in what i expect (feel free to correct me) is a code path for a little-used codec
maybe we'll just see them remove support for these long-tail formats the way linux has been removing drivers for similar reasons https://www.phoronix.com/news/Linux-Retiring-Moxa-Driver
it's widely used but in "industry" applications. so ffmpeg is probably being used in a lot of offices (studios) and maybe even being included in end user software.
Understatement of the year. Almost everything that processes video uses ffmpeg.
The README for the fuzzer is an AI slop mess. https://github.com/daedalus/fuzzer/
OP here: You are welcome to send a PR if you like. I'll be grateful if someone makes the readme more human.
I imagine the discussion will center around this application of AI, but to me this is just the Nth proof of the proven fact that you must build ffmpeg, if you insist on using it, with only an allow-list of file formats that you expect to encounter, and not with the kitchen sink of stuff you are never going to need.
send patches
...they did.
Where?
They only suggested a basic guard, chich can be useless if this case never happens
Generating an incorrect input file seems to be the easiest task of all for any fuzzer.
Generating correct input to get deep into the call stack and then finding something is the hard part.
[dead]
[flagged]
Unrelated to the submitted link -- just checked your comment history and all of your comments are AI-generated like this one. What's the motivation for this?
[flagged]
Nice find. The interesting part isn't "AI wrote the fuzzer." It's that a cheap random harness still hits classical bugs in ancient parsers. Keep the corpus; throw away the hype.
The fruits of using LLMs to code. You'll waste far more time finding what it quietly and subtly wrecked than you would have if you just coded it yourself.
Those sneaky LLMs going 7 years into the past and committing as a human:
https://code.ffmpeg.org/FFmpeg/FFmpeg/commit/8eda3c7f91e1a5b...
It’s obviously Claude 69 with time travel functionality, that’s too dangerous to release to public. They’re working on space-time limiting sandbox to prevent these issues.
You're not reading it right. The bug was found using a vibecoded fuzzer.
I wonder from where Claude stole this fuzzer.
Or it used something called an "analogy" which is a valid way to solve new problems.
The very first line of the bug report:
> This is a bug found with our fuzzer: https://github.com/daedalus/fuzzer/