🐍 ❤️ 🦀
Published October 14, 2022
This video features Markus Holtermann at DjangoCon Europe 2026 in Athens, Greece.
TALK: Oh, I Found a Security Issue by Markus Holtermann
This talk is your behind the scenes guide to Django's best in class security processes. I’ll give an introduction to how the team handles security issues: the triaging, fixing, disclosure process, and releases. I will then review the history of Django’s security issues to identify hotspots and areas to look out for. Lastly, I will explore the impact of AI and LLMs on the security of Django as well as its security team. The talk will give you everything you need, to help you interact with Django’s security team when needed, and show how Django's security process can act as an example for other open source projects.
More information: https://2026.djangocon.eu/speakers/BWWWMK/
Follow DjangoCon Europe on social media:
*X - https://x.com/DjangoConEurope/
*LinkedIn - https://www.linkedin.com/company/djangocon-europe/
*Mastodon - https://fosstodon.org/@djangoconeurope
#django #djangocon #djangoconeurope #djangoconeurope2026
Automatically transcribed, so expect mistakes in names and technical terms.
So
Speaker 1: Markus Holdermann is our next speaker. He's the member of the Django security team for many years. So He will talk to us about security. It is a follow-up of a talk uh that he gave a few years ago. But now with all of the things that are happening in the world, we need a follow up. So again, please give a warm welcome to Markus.
Speaker 2: Thank you. Yeah, good morning. Oh, this is loud. Good morning everybody. Um thank you for the organizers for running DjangoCon yet another year. I'm really looking forward that this is in Athens, um yet another country on the list of countries where DjangoCon has happened Thank you, Carlton, for that very inspiring and very interesting talk in this morning to kick off DjangoCon because I think there's a lot of thoughts that we can explore for the years to come and continue Django's presence in the market in that sense. Um yeah for those of you who don't know me, my name is Markus Holterman. I in my day job I work as at Kraken.
Speaker 2: We build a software to essentially provide energy suppliers. with a tool to sell green energy better and more efficiently. And I'm also a member of the Django security team and that's my role in which I'm standing here today and talk about Django security. Marius just said I gave this talk years ago, yes, 2017, 18 in some form or fashion, and I figured there is a reason to pull this out of the cupboard again and do a variation of it. So bits and pieces might be still well might be familiar to some of you. Um but a lot of things have changed actually.
Speaker 2: I want to start off with a brief of overview of Django security policy, how we deal with security issues in Django, what's What happens if something is being reported? I wanna go into how you can maybe learn from that process and take that on for your own projects. And last but not least, I want to look a bit into the history of security issues in Django, but also what's happening right now and where the journey might be going in the future. Can I get a show of hands who have seen an email such like that? That is not enough Um counter qu okay, it's the first
Speaker 2: day counter question. Who has not seen that? Everybody else is sleeping? Okay. Let's look into what it takes for the fellows Natalia, Jacob, Sarah, or anybody who has come before them to actually send out such an email. What it takes is that is written in down in the Django security policy. It defines how security issues should be reported, the do's and don'ts, the things that As we as Django security team want you as potential reporters or anybody out there to follow and adhere to to make our lives easier. It explains how the secure team evaluates the reports, what
Speaker 2: severity we decide on, and how we decide on the severity based on the reported issues. And what we is the variety we assign there. And it anslines which Django versions will receive the fix. And then also how the issues are eventually disclosed. I'll dive into the do's and don'ts of reporting in a in a moment, um, but the rest is rather straightforward. So for the versions, it's up to the last four versions. I mean we'll fix the problem on the main branch if it's in the main and hasn't been Fixed for some other reason there anyway. And then the last two major releases. Right now that's 6.
Speaker 2: 0 and 5. 2. And then also the last two LTSs. Which is five point two and six point uh four point two. And that's said, four point two got the last stable release beginning of this month. If you're still running on 4. 2, now is the time to upgrade to 5. 2 at least. If you don't, you will not get any security fixes. There will not be another release. You really should be updating now Before the next month, potentially, maybe. Um yeah, as for the severity levels, there are three of them: low, moderate, and high Um they have coup certain certain issues we typically classify in with certain severity
Speaker 2: because typically they are either critical or high or low, but technically we can change the severity per issue because your DOS factor might actually be so easily to exploit that low is not the right choice. Great. Now let's say we've you found a security issue. Or suspect you found one. How would you report that issue to us? Anyone got an idea? To where? Wonderful, exactly. Um it's the at least right now
Speaker 2: preferred way, the thing that we typically use, and that's also the um email uh group uh mailing list group that we use for the day-to-day communication. We also have Hacker One as a tool too for reportings. We also am trying out the GitHub security vulnerabilities feature with some folks who wanna who report things. Um but generally if you send an email there that's probably your best choice. Um When you do so, you should give us the minimum information necessary. Please don't write walls of text, because that's probably contains too much information and hides the important part between the text.
Speaker 2: Include a realistic proof of concept. Don't come up with something that is not going to be there in any regular project. And if you have the time, we welcome a patch. If you don't, that's absolutely fine. We can write the patch um and do a bit uh a bit of back and forth with you as a reporter to make sure the patch is correct and makes sense and fixes the real problem. But yeah, please don't construct arbitrary edge cases or report issues that go against some of the fundamentals of modern web development. If you d take your raw user input and just pipe it to I don't know where, and you say, well now I have a remote code execution.
Speaker 2: Yeah, you should be validating your user input before. What follows is we essentially do a triage. We take your email, we read through that, we figure out what is the issue, what is the core of that, how do we fix this? Um members of the team, typically the fellows, will respond to your email with either, yep, that seems like a valid report, uh, we'll be in touch for further doing the investigation. They might respond with well we'll need to look into that. It can take a while until we get back to you. But also, yeah, what you submitted isn't really an issue Um personally, I'm probably not gonna respond to any or not be involved in topics on the SQL injection area
Speaker 2: because that's Then goes down into the ORM parts where I'm personally not too familiar with, but stuff on the HTTP, HTML part um HTML par um rendering or HTML parsing uh HTTP parsing, that's something where I probably would would be involved also in fixing. But essentially that's up to anybody on the security team team to decide if they want to contribute or have time to contribute on a specific issue. Assuming there's a legitimate issue, we and we have a fix, we'll send that, as I mentioned, back to the reporter and see and ask them if they can confirm that it fixes the problem that they reported. If it does, well great, we continue.
Speaker 2: If it doesn't, well, we go back to the drawing board and figure out how to fix this. Assuming the patch is correct, we'll continue issuing with a common vulnerability and exposure number, a CVE. And since October 2025, the DSF is its own CNA, so CVE numbering authority. That's the authority or authorities that are allowed to issue CVEs for certain projects. Back in the days, that's what only MITRE or whatever they're called in the States. It was a think kind of federal agency, something something. And then you send them an email like, hey, we have this issue in this open source project, blah blah blah blah blah, and they'll get back to you with
Speaker 2: please use this CVE. And they kind of distributed that onto more shoulders, bigger open source projects, companies can become um numbering authorities for their products. It also means they are now the owners of those CVEs. Nobody else is able to issue CVEs for those products. Back in the days, you could have gone and sent the request for a CDE for whatever product. Hey, I want a CDE for this product, for this bug. They would issue that. And that company would have no way to actually really deal with that. They could try to argue against that, but realistically, not really.
Speaker 2: Well, once we got the CVE, we gotta send out an email to the announcement list. There's going to be a security release in a week. And on the day of the release, the patch or the patches will be applied to the uh repository, and we will issue the release. And then following that, we announce that on the web on the blog and copy the link for that into the forum, into the mailing list, into the email that you saw before. Now, how does appl how can you apply that to your own project? First and foremost, communicate. Communication is the key in this entire process.
Speaker 2: Say how people can report something to you and then go and actually monitor how they um how what they report. There's no point in you putting out an email address where they can report something and then you not really respond there. But then you also as an open source developer on a soul on a sole project, you don't need to be online 24-7 for that. Tell people, well, I'll respond to you typically within a week. That's more than enough time. Or two weeks, whatever. Whatever makes sense to you. After all, you're doing that in your free time properly, and please don't burn out over that And then find a way to fix the bug.
Speaker 2: Again, communicate with the reporter, come up with a fix, either by yourself, have ask the reporter to help you out if they have the bar time, if they want to contribute. And when you have the fix, check with them before you make the fix public. Because there's no point in doing a security process if you try to fix this whole thing in the public anyway. And then commit, release, and inform people. Again, figure out a way how you want to communicate that. If you're on GitHub using GitHub security advisories might not be the worst idea. They're not too shabby. Um and they're fairly easy to set up. And then for your own sake, accept the fact that you had a security issue.
Speaker 2: Django had over 140 over its years. We all make mistakes, so will you, so will I, and that's just the cause of nature. Right. So let's look a bit into the history of security issues in Django. Before we want to look into the future, which you might have an idea what this could be, let's look into the numbers of and some statistics. Here's the numbers of CVEs per uh for Django per year. I mean it's a statistic you can believe me or not. This is the data I pulled from the Django
Speaker 2: security um archive. And I hope I th I didn't screw up my script to pull the data there. So over the last sixteen years since when we started We've consistently had around five to fourteen valid security reports per year. Except for 2017. I have no clue what happened there. And then for 2026 we also have like 13 f um I think yeah 13 reports or valid security issues. It's April I'm not sure where this is going. Quite unsurprisingly, these are the um
Speaker 2: CBEs per type of issue, per classification. denial of service is at the forefront that's just too easy to make a mistake there and on the other hand uh oftentimes they're not really an issue these days until you have the f until we have the fix in Django or until you have the time to actually update to Django to a fixed version. The solution sometimes might be okay I'll throw another two, three VMs there or increase the memory for for the time being and you'll be fine. But that doesn't mean it's always the case. Going back to the this DOS
Speaker 2: might actually be high here, not low. Then SQL injections and information information disclosure issues are up there as well as is XSS, like cross-set scripting. After all, it's a web framework. The other issues occur rather seldom. Django seems to be pretty good at not making the issues in the first place, which is good, I guess. And then we can look into the four top ones and see how those evolved over the years. I mean, again, denial of service remains at the form from beginning 2024. SQL injection clearly became more, while XSS kind of subsided in
Speaker 2: uh for the last three and a half years. And as for information disclosure, well they occur like once, twice a year, but We seem to be making mistakes there, but it doesn't seem to be like particularly terrible, I think. So what we've looked at are the public issue published issues. And I'm also promised to talk about SQL injection and inform information disclosure issues. Let's look at the reported issues, like the statistic on As I mentioned earlier, um and as you previously said Um security at Django Project. com is the preferred
Speaker 2: way to report that. But we have also Hacker One. And going through years and years of security archive was kind of so much work that I gave up on that and then went over to Hacker One and actually just looked at Hacker1. These are the Hacker One submissions since we have that account since 2016. As you can clearly see, reports skyrocketed in 2025. Going from under 20 to almost 70. That's just reports. And with that, the quality of reports plummeted. As you can see in the pink line, the ratio of valid versus invalid reports. It's the number of
Speaker 2: or percentage of invalid reports. compared to the valid ones. How that did that happen? Anybody got an idea? No. Everybody giggling? Nobody raising their hand? Thank you. Um exactly. Most of the reports appear to be AI AI written or at least written with the help of AI. And um The quality of them, of some of them especially, is abysmal. Let's look at an example. Published November 5th, 2025. Potential SQL injection via underscore connector keyword
Speaker 2: argument in query set and queue objects. The issue was originally reported middle of September, so within six weeks she we actually fixed uh went from report to release. The problem though is that not that specific the issue itself. Problem is what came after. Since the announcement, we've received a dozen reports on the same thing or a very similar thing or targeting the underscore connector attribute in one way or another. Guess what? We closed all of them as invalid. Some because they were just plain invalid, because they were lacking best practices, they were fabricating code.
Speaker 2: Or using not just didn't validate user input. Like that's a dozen reports in that statistic, that's just waste of time. Because we need to go through all of the reports. valid or invalid reports we've got so far in twenty twenty six. I mean the statistics is grayed out for a reason there for twenty twenty six. Any guesses? Okay, so we got ten valid reports and fourteen invalid ones.
Speaker 2: That is fifty-eight percent invalid. The numbers for twenty twenty-six so far looks w look worse than twenty twenty-five or twenty twenty six, uh twenty two twenty twenty-four. The ratio of valid versus invalid ports is almost back to where we were in 2024. But at yeah, at the end of the first quarter. We are at more than a third of reports in tot compared to 2025. That means if we extrapolate, we end up at 80-90 reports. via hacker one we still have emails they are not emails are not in that statistic that's exhaustive
Speaker 2: that's annoying I gotta admit when I put the statistics a month ago first, the ratio of valid versus invalid was at 65 %. We are now at 58. So That's better I guess. Maybe L LLMs turned became better. I don't know. Um I don't think what's the new thing from from um Anthropic chord mythos or something? I don't think that's playing a role here. This is just people Or LLMs being probably better, I guess. So what does the future
Speaker 2: look like that's ahead of us? Well, the Internet Bug Bounty Program, that's the the system online that reporters could go to and claim a bounty for a security issue that they found in Django. They stopped taking submissions on March twenty seventh this year. Quoting The discovery landscape is changing. AI-assisted research is expanding vulnerability discovery across the ecosystem, increasing both coverage and speed. The balance between findings and remediation capacity in open source has substantially shifted.
Speaker 2: The system that was there to pay out bounties for open source projects, which was funded through Big IT mostly big IT companies said we don't have a way to pay this. Or pay out those bounties essentially because of how the landscape is changing. Django security policy already states that AI assisted reports must disclose AI tools and verify issues before reporting. And avoid fabricated code or references to non-existing features in Django. Yes, we've seen reports saying there is this bug in Django, referring to code in Django that just doesn't exist in Django.
Speaker 2: What the heck? In another example, we s we received four reports within four minutes from the same reporter. Turns out the reporter is an intelligent security platform combining advanced AI and domain expertise. We've closed two of those issues as invalid. Guess fifty percent chance of being right They could only f provide fabricated code that avoids some of the best practices of application development. There have also been reports with numbers increasing recently about issues relying on calling Django's private API.
Speaker 2: Last time I've that I've checked, that's out of scope for the security policy. But AIs and LLMs simply do not care. As a result, we've now put a few more measures in place to handle the flush of reports. I think last week, two weeks ago, we made a change to the policy that only one there must only be one report per person ongoing. Additionally, if you use report or submit reports with overly long text, walls of text We're just not gonna care and not gonna look at them unless we figure well this seems legit or until somebody wants to spend time on that.
Speaker 2: But we don't need to anymore, because we clearly state you need to be precise and concise. And if you don't, this is a waste of our time. And another thing on Hacker One, you can say you close this report as informational, as invalid, or as PAM. Typically we went with invalid, which gives a reputation of minus five. Maybe we just need to go with minus ten. And for repeated people um submissions by the same person maybe we need to be a bit more strict and go and block and ban them That's it. Thank you.
Speaker 1: Thank you. I think we have time for one or two questions
Speaker 3: Thank you for the um overview, Marcus. Super. Um how does the security team go for capacity in the face of this wave of
Speaker 2: So a lot of work is done by the fellows these days and arguably too much time spent by them on security issues and not spending on other things. I'm not sure how much it would help having another one, two, five people on board. Maybe, maybe not. I think the problem right now is more that we need to find a way of working in a more structured appro um using a more structured approach. As I mentioned earlier, the security email address is the one where we discuss things. Mostly because We 've and Django has always been a um a project where security reports have been discussed completely in the dark
Speaker 2: and then once There is a re um a fix for that that is being dis um published, which means the issue tracker that we currently have track. it doesn't support private issues. So keeping track of all the reports currently becomes a bit of a pain. And we're looking into options how we can deal with that. Like all the options that are kind of on the table kind somewhat suck in a different ways. Hacker one requires your users to agree to their terms and services. GitHub security advisories don't really support some of the processes that we have or want to have. Emails don't give you a shared labeling, for example.
Speaker 2: Like It is hard. Um so adding five people to the team doesn't solve the foundational process problems, I think. And that's what we need to figure out first.
Speaker 1: Anyone not from the first row, maybe?
Speaker 4: Hi. Hi. That SQL injections are getting more and more prevalent. The first security issue that I learned about at age of twelve I think was the SQL injection. How come it's still a thing? Do you have any insight to that?
Speaker 2: Yes. Um So taking that example that I also had on here, this one, I think if I'm not mistaken, that was a SQL injection as well. If you twist things the right way and start use to use undocumented kind of private API, but maybe not because it's somewhere exposed and you maybe actually be using that realistically. and then use somewhat validated user data with somewhat legitimate inputs and pipe it in there, then you have an attack vector. Arguably in hindsight may I
Speaker 2: This may or may not have been sick might have not been a security issue. We said it is one back in the time, so we stand by that. Um but uh There is the way the RM works and the pieces in Django that r turn things into SQL They need to support both the user input, but they also need to s uh support um the the syntactical the the SQL syntax Strings just say to say. So if you the connector here is the thing that essentially in the query set Gilson says is it or or an and in the condition? And If you overwrite that with
Speaker 2: simply put, with a string that is neither and nor or, but with quotes comment blah blah blah like your typical SQL injection so to say, then you have a SQL injection. And Django can't really escape that because it needs to be either one of those strings. And so there are s legitimate ways to create SQL injections. If you twist your eyes properly and turn the data properly. But for the normal cases, like your normal filter, your normal annotation It really shouldn't.
Speaker 1: Okay, the last question. Someone at the top. Thank you.
Speaker 5: Yeah, so thank you for the talk. Uh I just had a quick question um you mentioned and showed up a bit about the uh misconfiguration and you talked about the issue with the uh False positives, for example. And I was curious if you think that maybe adding like additional guards against uh misconfigurations might help. reduce the likelihood of for example LLMs making false positives or sending like invalid reports.
Speaker 2: So if cur if LLMs start fabricating code and say this is an issue in Django where code just doesn't exist, then I'm not sure how much having a guard that prevents a SQL injection in sudden in certain place would help there. Um
Speaker 5: yeah but I meant for example also like um uh let's say something that that is just a misconfiguration on the user side like a CSRF issue or like you could implement guards against that in Django to enforce um stricter logic at kind of a detriment to the user experience maybe. I'm just curious if that would uh
Speaker 2: I'm most of the or maybe all of the feature security features in Django have a deliberate option to turn them off. There is reasons why you or there are cases where you cannot have a CRS off on on a view, for example. If you know for a fact this is not something that's can be used in certain ways, for example. If you create such a view and then report, well, there is an issue now. Then fundamentally I I as a developer I don't want to have CRSF here on on this on this view. But if I then go and say now it's subject to C RSF attacks, then
Speaker 2: What is the point having a You see the problem? Like if we go and say you cannot turn off C SRF Then people would come and say, well, I have legitimate use cases where I cannot use CSRF. Like both sides of the table need to work. And that's the problem that that we need to uh yeah deal with. Query set extra allows you to insert raw SQL. If you know what you're doing, if you properly escape parameters, this is not a problem. If you just pipe unvalidated user input there, well sure, you're gonna shoot yourself in the foot. But then you probably maybe possibly shouldn't be using raw SQL
Speaker 2: there in the first place or at least make bloody sure that you validate your user input. If you don't, then I actually didn't see uh didn't count how many reports there are for well extra allows SQL injection. Probably too many of them.
April 15 - 17, 2026
Athens, Greece
Note: We understand that names change, people change, and bodies change. We respect each individual's journey and privacy. If you have any concerns about a video or need us to remove content, please don't hesitate to contact us. We will handle your request with care and promptly address any issues.
Published October 26, 2026
Published October 26, 2026
Published October 26, 2026
Published October 26, 2026
Published October 26, 2026
Published October 26, 2026