cybersecurity / artificial intelligence / ai agents / software engineering
AI Hacker Attack on My App: 10 Fake Accounts in 4 Hours, No Data Lost
On the evening of November 20, someone probed my trading platform for about four hours, and the pattern of their requests suggests an AI agent was doing the work. It started at 6:12 PM with error alerts and forged login tokens. By 6:53 they had created an account and bypassed email confirmation through a hole in my app. By 7:08 they had posted comments on my blog that tried to inject code into the page, and my sanitizer removed it. I banned accounts and blocked IP addresses, and they came back from Japan and then Singapore, going after user data, my admin tools and my cloud secrets. None of those attempts worked. By 9:55 PM I had banned their tenth account and shipped my final fix, and they have not sent a request since.
No user data was affected, and I fixed 13 security holes before sunrise with Claude Code. The requests match ARTEX, an open-source penetration-testing framework that the FBI issued an alert about one week earlier.
The video walks through the timeline with the real alerts, comments and IP addresses. It also explains why the experience changed my view of open-source AI models, and why I think the closed labs should let an owner attack-test their own system.
Transcript
0:00Someone from Israel tried to hack my trading platform The scary part is it wasn't like a noob hacker just probing endpoints
0:07They were assisted by the use of AI And my opinion on open-source
0:12AI models has now forever changed It started at 6:12 PM where I started to get some error alerts that people were failing to log in
0:21I didn't think much of it at first because it happens all the time
0:26But by 6:30 PM my phone was blowing up and I started to see what all the noise was about
0:32And what I saw wasn't someone failing to log in
0:34It was someone trying to forge a login token I used
0:38Claude Code and I started to mount my defense By 6:53 PM they had managed to create an account even bypassing the mandatory account confirmation step
0:48I saw that my app had a security hole The six digit confirmation code was visible through the server
0:55By 7:08 PM they had posted eight different comments on different blog posts
0:59And each of those comments had the sole purpose of trying to submit code in the browser
1:05They tried to exploit a cross-site scripting vulnerability attack
1:09Fortunately my system held And my sanitizers removed any hints of malicious code
1:14At 7:20 PM I banned their account and I banned their
1:18IP address And I thought that would be the end I used
1:21Claude Code and started to patch all of my security vulnerabilities
1:25But that wasn't the end In fact it had barely just started
1:29By 8 o'clock they started to flood me Flooding all my contact forms I blocked them at the
1:35Cloudflare edge layer So they went straight through it attacking my server directly
1:39So I locked the server down So they can only be accessible through
1:43Cloudflare And I thought that would be enough I was wrong
1:46By 9:03 PM they showed up again This time with an account from
1:50Japan And now they were angry They were trying to exfiltrate data for my users
1:57It didn't work So I banned them and they came back again
1:59This time through Singapore They started to attack my admin tools
2:03Getting admin access to the API They also probed my
2:08AI model hoping it would just give away my cloud secrets
2:12It didn't work By 9:55 PM I had banned their 10th account that they created
2:16And I pushed my final change to harden my defenses
2:19And they didn't submit another request since Here's the part that should scare you
2:24This wasn't a noob just walking down a checklist slowly just probing endpoints
2:28They were absolutely aided by a language model From the patterns of the requests it seems they were aided by
2:35ARTEX An open-source software that's designed for penetration testing
2:40And it was already identified by the FBI for aiding in cybercrime
2:45Combined with an open-source abliterated model Like DeepSeek v4.1
2:50Flash Uncensored The framework would allow you to do anything
2:54Including trying to do an attack No user data was affected
3:00And I fixed 13 different security holes before the sun rose
3:05It was like those hacker movies Where their hacker bypassed security
3:09And I had to mount my defenses And I had to do this that and the third
3:12And it was crazy Now here's the one part that irritates me the most
3:17After everything was said and done I wanted to use
3:20Claude Code to secure my own system And Claude wouldn't work
3:25Codex also wouldn't work Wouldn't perform the attack So I'm now left open for someone to try again
3:33The only way I can defend myself is if I use these uncensored models too
3:37That doesn't make any sense This entire experience changed my perception on open-source
3:42AI models If you're hosting an AI model that can perform cyber crimes
3:49You should be held accountable for what people do with it
3:53Like back in the day this type of hacking actually required work
3:57And now you can point an AI and it can do it by itself
4:01This doesn't make sense And the closed source labs they're not off the hook either
4:06What do you mean I can't try to hack myself
4:10That's outrageous So if you're building software and especially if you're a vibe coder
4:14You don't know about security principles You have to be careful
4:19Take some courses Use AI to harden your defenses And don't just assume they can't happen to you
4:25Because I didn't think it would happen to me either
4:29Stay safe out there
Join the conversation
Loading conversation…