With a concrete bug report like “using codec xyz and input file f3 10 4d 26 f5 0a a1 7e cd 3a 41 6c 36 66 21 d8… ffmpeg crashes with an oob memory error”, it’s pretty simple to confirm that such a crash happens
Google’s big sleep was pretty good, it gave a python program that generated an invalid file. It looked plausible, and it was a real issue. The problem is that literally every other generative AI bug report also looks equally as plausible. As I mentioned before, curl is having a similar issue.
Stenberg said the amount of time it takes project maintainers to triage each AI-assisted vulnerability report made via HackerOne, only for them to be deemed invalid, is tantamount to a DDoS attack on the project.
So you can claim testing may be simple, but it looks like that isn’t the case. I would say one of the problems is that all these people are volunteers, so they probably have a very, very limited set of time to spend on these projects.
This was the first search hit about ffmpeg cve’s, from June 2024 so not about the current incident. It lists four CVE’s, three of them memory errors (buffer overflow, use-after-free), and one off-by-one error. The class of errors in the first three is supposedly completely eliminated by Rust.
FFMpeg is not just C code, but also large portions of handwritten, ultra optimized assembly code (per architecture, too…). You are free to rewrite it in rust if you so desire, but I stated it above and will state it again: ffmpeg made the tradeoff of performance for security. Rust currently isn’t as performant as optimized C code, and I highly doubt that even unsafe rust can beat hand optimized assembly — C can’t, anyways.
(Google and many big tech companies like ultra performant projects because performance equals power savings equals costs savings at scale. But this means weaker security when it comes to projects like ffmpeg…)
Have any of the google-submitted vulnerability reports turned out to be invalid? Project Zero was pretty well regarded.
Yes I know about the asm code in ffmpeg though IDK if it’s doing anything that could lead to a use after free error. I can understand if an OOB reference happens in the asm code since codecs are full of lookup tables and have to jump around inside video frames for motion estimation, but I’d hope no dynamic memory allocation is happening there. Instead it would be more like a GPU kernel. But, I haven’t examined any of it.
Anyway there’s a big difference between submitting concrete input data that causes an observable crash, and sending a pile of useless spew from a static analyzer and saying “here, have fun”. The Curl guy was getting a lot of absolute crap submitted as reports.
From the GCC manual “bug criteria” section:
If the compiler gets a fatal signal, for any input whatever, that is a compiler bug. Reliable compilers never crash.
Project zero was entirely humans though, no GenAI. Project big sleep has been reliable so far, but there is no real reason for ffmpeg developers to value project big sleeps 6.0 CVE’s over potentially real more critical CVEs. The problem is that Google’s security team would still be breathing down the necks of these developers and demanding fixes for the vulns they submitted, which is kinda BS when they aren’t chipping in at all.
Anyway there’s a big difference between submitting concrete input data that causes an observable crash, and sending a pile of useless spew from a static analyzer and saying “here, have fun”
Nah, the actually fake bug reports also often have fake “test cases”. That’s what makes the LLM generated bug reports so difficult to deal with.
6.0 is pretty serious according to the rubric. Are there some worse ones? Yes Google is acting obnoxious per your description. It makes no sense to me that they are balking about supplying some funds. They used to be fairly forthcoming with such support.
I can imagine a CI system for bug reports where you put in the test case and it gets run under the real software to confirm whether an error results, if one has been claimed. No error => invalid test case.
Google’s big sleep was pretty good, it gave a python program that generated an invalid file. It looked plausible, and it was a real issue. The problem is that literally every other generative AI bug report also looks equally as plausible. As I mentioned before, curl is having a similar issue.
And here’s what the lead maintainer of curl has to say:
So you can claim testing may be simple, but it looks like that isn’t the case. I would say one of the problems is that all these people are volunteers, so they probably have a very, very limited set of time to spend on these projects.
FFMpeg is not just C code, but also large portions of handwritten, ultra optimized assembly code (per architecture, too…). You are free to rewrite it in rust if you so desire, but I stated it above and will state it again: ffmpeg made the tradeoff of performance for security. Rust currently isn’t as performant as optimized C code, and I highly doubt that even unsafe rust can beat hand optimized assembly — C can’t, anyways.
(Google and many big tech companies like ultra performant projects because performance equals power savings equals costs savings at scale. But this means weaker security when it comes to projects like ffmpeg…)
Have any of the google-submitted vulnerability reports turned out to be invalid? Project Zero was pretty well regarded.
Yes I know about the asm code in ffmpeg though IDK if it’s doing anything that could lead to a use after free error. I can understand if an OOB reference happens in the asm code since codecs are full of lookup tables and have to jump around inside video frames for motion estimation, but I’d hope no dynamic memory allocation is happening there. Instead it would be more like a GPU kernel. But, I haven’t examined any of it.
Anyway there’s a big difference between submitting concrete input data that causes an observable crash, and sending a pile of useless spew from a static analyzer and saying “here, have fun”. The Curl guy was getting a lot of absolute crap submitted as reports.
From the GCC manual “bug criteria” section:
That sounds like timelessly good advice to me.
Project zero was entirely humans though, no GenAI. Project big sleep has been reliable so far, but there is no real reason for ffmpeg developers to value project big sleeps 6.0 CVE’s over potentially real more critical CVEs. The problem is that Google’s security team would still be breathing down the necks of these developers and demanding fixes for the vulns they submitted, which is kinda BS when they aren’t chipping in at all.
Nah, the actually fake bug reports also often have fake “test cases”. That’s what makes the LLM generated bug reports so difficult to deal with.
6.0 is pretty serious according to the rubric. Are there some worse ones? Yes Google is acting obnoxious per your description. It makes no sense to me that they are balking about supplying some funds. They used to be fairly forthcoming with such support.
I can imagine a CI system for bug reports where you put in the test case and it gets run under the real software to confirm whether an error results, if one has been claimed. No error => invalid test case.