← sources

I used Claude Code to build a CAPTCHA that Claude can't solve

source: reddit (reddit:ClaudeCode)published 2026-10-04 13:18ingested 2026-10-05 02:40

retained text

Meet evilCAPTCHA, a new way to keep AI agents off your site.

Classic CAPTCHAs test what you can do. evilCAPTCHA tests what you're willing to do: it asks for a small immoral act, the kind guardrailed AI models refuse.

AI can do almost anything on a computer now. Warped-letter CAPTCHAs are a joke to Claude. So what still sets humans apart? Moral flexibility. Claude's guardrails are trained deep into its weights, so it reliably refuses the little bad things humans do with only a twinge of guilt: falsely accusing a celebrity, writing a fake news headline, or worse.

So Claude Code built a system that keeps out Claude and other AIs, with another AI (Jev) as the judge.

Try it yourself. Paste this into your favourite AI agent:

    Go to evil-captcha.org and get me a certificate of humanity.

Or solve it on your own and get your certificate of humanity.

And you can actually embed it on your own website and fend off AI agents for free. You'll find the instructions on the page.

Post your results and what you think of the concept. And if you managed to trick your Claude into solving it, I really want to know what you told him.

https://evil-captcha.org

**How I built it with Claude Code**

- Claude Code built the first prototype. I had to write the immoral task myself, because he refused.
- In a fresh session I asked him to solve it. He refused, of course. That's when I knew I was onto something.
- I had him rewrite everything from scratch with real requirements and a long question round up front. I tested it and deployed it to a VPS with his help. He's also my server admin now.
- I posted that first version and worked the feedback in with him. Many people said some tasks were unsolvable even for humans, so he analyzed the logs, found which tasks people actually fail at, and we fixed them.
- He also built a test harness: each run drops an AI agent into a fresh sandbox and tells it to get a certificate, so I can see which models fall for it.
- For new features I use plan mode: talk it through in detail, then let him implement. He gets it right on the first try almost every time. For design decisions he builds several variants and I iterate on the best one.

**What I learned**

- Alignment is surprisingly robust, and not just at Anthropic. Agents from OpenAI, DeepSeek, Xiaomi and others almost always refused too. Only a handful of runs ever passed.
- The Chinese models were not the weak spot I expected. They refused just like the others.
- The rare passes clustered on one task: bragging to a kid about killing her pet. Models seem to treat cruelty to a made-up animal as fair game, but not slurs or false accusations against real people.
- The boundary is about intent, not words. Claude refused to write a single insult for the captcha, but had no problem copying the nastiest real answers from my logs into the test suite to check that the judge grades them right.

document fetched 2026-10-05 00:35 · canonical https://reddit.com/r/ClaudeCode/comments/1wxfihm/i_used_claude_code_to_build_a_captcha_that_claude