返回资讯中心
外部精选
软件工程
#开发者实时流

Scammers, exit nodes, and my part-time job fighting abuse

Code and prose written on val.townScammers, exit nodes, and my part-time job fighting abuse

Hacker News 最新jackyalcine2026/10/17 · 2026/10/1724 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

Code and prose written on val.townScammers, exit nodes, and my part-time job fighting abuse

Tom MacWrightOct 16, 2026

Unfortunately, spammers and scammers are constantly trying to use Val Town as part of their toolset. We had a honeymoon period of a year or so when it was only some enthusiasts using the platform. But then we appeared on their radar. Ever since then, I've had the part-time job of battling the Indonesian gambling websites, OpenAI token resellers, rootkit distributors, and all other kinds of scoundrels.

It can be tough work, not just technically but mentally and emotionally, and it's given me a deep appreciation for the people doing harder work and more of it at the Trust & Safety teams in big social networks. After spending some time doing abuse prevention, I usually need a long walk and some time to paint watercolors or play the guitar. A creative reset.

But that's not to say it's uninteresting. It's interesting as heck. Since I haven't written about it before, here's what I've learned about moderation, abuse prevention, spam, and so on.

I was shocked to learn how many big platforms are part of these attacks. For example, we have long had a problem with people hosting phishing attacks on Val Town, in particular fake cPanel or webmail login pages. But even though Val Town lets you host databases, backend applications, send emails, and more, their code would just be a static HTML page that would send the stolen credentials to one server, load JavaScript from another one, and so on. And they'd use some other stolen resource to send the emails with links to that bogus login page.

Obviously I'm not rooting for them to host the whole thing on Val Town. By distributing the parts across services, it surely becomes a little harder to track it all down and identify that one of the services is involved. And it's consistently shocking which services are involved - JavaScript from a storage service hosted by a major publicly-traded company, a database by a major startup, files in popular file-storage sites. I even encountered one attack that used the Internet Archive as a link in the chain.

One lesson we all keep learning is that every application can be distilled down to its component parts and used for something unexpected. An NPR comments section can be a chatroom for teens. GitHub Actions can be used to mine cryptocurrency. YouTube can be infinite file storage. Similarly, any website that lets you host a file is an abuse vector.

A fake login page for a real bank is an easy red flag for bad behavior, so stealing bank credentials is a good rule for a detector. But then you identify someone who's transmitting the session cookies for an online bank between browsers and, after learning about them and their use-case and the fact that it's a purely personal tool, that actually turns out to be legitimate. The goal is to find enough bad guys to protect Val Town's services and domain and to crack down on the internet-wide problem of abusive behavior, but trying to find every malicious actor is probably an impossible feat. My favorite columnist, Matt Levine, wrote about this phenomenon with banks recently - they also need to find most of the bad actors, but they routinely miss some and that job then falls to regulators.

I find myself in a tension between natural curiosity about why these people are doing these particular schemes, and the practical principle that it doesn't really matter that much: scammers aren't really putting that much effort or intentionality into setting up these systems because they aren't very high-reward anyway. They have to expect to get shut down at some point, so they're not like engineers crafting some beautiful system.

So our categories grew to be more broad in terms of topic but narrow in terms of functionality: are we seeing telltale obfuscated code patterns, or people sending a password to some other service? Is it gambling or crypto-related?

Anyway, why Indonesian gambling websites? As far as I've been able to learn, the country has both skyrocketing interest in gambling but also a total ban on gambling, which they're attempting to enforce. Multiply that by a population of 287 million, and you end up with a lot of operators trying to host illegal websites anywhere they can. They're using Val Town to try and get SEO or just run the gambling apps as long as they can.

Here's what we actually see, in terms of what gets accounts banned:

The long tail includes things like attempted SSRF exploits on Val Town itself, and people trying to distribute other hacking tools.

When I started building the system for abuse prevention, it relied on actual hard deletion of user accounts. Wiping accounts was rare and a simple system was enough. But once we introduced automatic detection, suddenly there are other risks: what if a trusted user trips the system or we need to audit the code that we just deleted? What we have is by no means a self-improving system, but banned user content is pretty useful for improving the LLM-based rules.

So I built a more advanced system, that supported silent banning, sending notifications, unbanning, and could store the reasons for why a particular account was banned. Now when accounts are clearly connected in the same network, the ban system can communicate that information. And when a user is automatically banned, they can get a notification and then contest and reverse the decision if it's wrong.

The first naïve system was based on keyword matching. There were a handful of predictable patterns and that worked.

But we quickly outgrew that and started using Netcraft as a primary signal. When a Val is saved, any web endpoints are sent to them for their audit, and if Netcraft reports something as abusive, we'll automatically ban. Netcraft has been incredibly accurate, even for sites with pretty sophisticated attacks. They don't catch everything, but what they do catch has been reliably correct.

But that only covers HTTP endpoints, and catches problems after they are online. Vals can run on intervals, send emails, and receive emails. And Val source code is all served as JavaScript modules on esm.town, so we could inadvertently serve as a link in a larger exploit. We don't want that either. So we additionally run all saved code through a cheap, lightweight LLM with tuned prompting. That produces both a score and a reason, and scores above a certain level get automatically banned.

And that only covers vals! Any textbox, any link on the internet can be part of spam. So that bot also crawls usernames, social media links, and bio information to catch people trying to use Val Town for SEO farming operations. An SEO farming operation usually looks like someone setting up a lot of accounts and creating accounts that advertise some online gambling, crypto, or AI service, and their "social link" in their biography will be to the service.

Those SEO schemes wouldn't work anyway: user-generated links on Val Town have always had rel=nofollow, so smart search engines disregard them for ranking purposes. But as I mentioned earlier, spam is a low-intent volume business.

We've also used VirusTotal as a secondary source of detection data alongside Netcraft. It's useful for different kinds of work, like manually diagnosing coordinated spam systems.

The problem with people hosting phishing websites on your subdomain is that they'll send emails with links to those fake login pages, and those emails will get flagged as phishing, and the links in those emails will also get flagged as phishing websites. So Google might act on that report, and the email deliverability of domains similar to the domain of the page will get docked. Also, it's wrong.

Chrome will flag all of the sites under the subdomain as unsafe sites, and you'll have to delete the content and contest the ban. Even if those subdomains are on the Public Suffix List, which kind of disclaims ownership of *.web.val.run, we're still really responsible and those sites are connected in the eyes of Google, and they probably should be.

Personally I think it's right to expect companies that are providing any kind of hosting of content to be responsible for that content. And even though there are some mild legal distinctions between the host and the hostee, the de-facto self-policing of the internet works, kind of. It could definitely work better, though.

In May this year, we got an extraordinary number of signups, all deploying identical code. Any guesses?

The internet blackout in Iran led to hundreds of thousands of people trying to get around a massive, restrictive firewall, and Val Town was one of the easiest places to deploy code. Unfortunately, hosting thousands of exit nodes is not our business, and it did a number on our servers, bandwidth costs, and IP reputation.

Other providers, like Deno, have solved this problem by restricting outbound traffic from new free accounts. We could do the same in the future, or try and find other fingerprints of generic proxies. Right now, we just try to ban users who are simply deploying generic proxies: it's a valid need, but Val Town isn't the place to do it.

Probably not. Definitionally, it's a moving target: today's crypto miners are tomorrow's stolen AI-token miners. There'll always be a new thing, and new schemes to get around the rules. This is why companies have whole departments devoted to this problem.

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →