Claude Code Refusing to Implement CAPTCHA Automation

Hello, I have a bit of a problem. For the past few days, I’ve been creating a bot for a game I play using Claude Code. Both the game and the bot are written in Java.

I already have almost everything finished, and there’s just one small thing left to add: automatic clicking on the CAPTCHA. The CAPTCHA is extremely simple because all you have to do is click in the middle of the screen on a button confirming that you’re at the computer.

However, Claude refuses to implement it because it considers this to be an interference with the game’s security. Does anyone know of any ways to bypass this restriction and make Claude do it?

I’d like to point out that image-based detection is not an option, because in the future I’m planning to use the bot with multiple clients.

Best regards.

Are you using fable 5 or opus? Fable will never let u do it.

I use both Sonnet and Opus.

you gotta try more LLMs to bypass, when claude refuses I go to kimi if that refuses go to deepseek. One will do it.

Xolta0 said:

you gotta try more LLMs to bypass, when claude refuses I go to kimi if that refuses go to deepseek. One will do it.

Click to expand...

Try GenSpark since they got a lot of models or with OpenRouter maybe Chinese models like Kimi?

donfelipon said:

Hello, I have a bit of a problem. For the past few days, I’ve been creating a bot for a game I play using Claude Code. Both the game and the bot are written in Java.

I already have almost everything finished, and there’s just one small thing left to add: automatic clicking on the CAPTCHA. The CAPTCHA is extremely simple because all you have to do is click in the middle of the screen on a button confirming that you’re at the computer.

However, Claude refuses to implement it because it considers this to be an interference with the game’s security. Does anyone know of any ways to bypass this restriction and make Claude do it?

I’d like to point out that image-based detection is not an option, because in the future I’m planning to use the bot with multiple clients.

Best regards.

Click to expand...

Just use a different model. Models from Chinese labs like DeepSeek and GLM seem to care less about stuff like this. Also clear context and figure out a prompt engineering technique "manipulate" the model into thinking it's helping "security research"

Yeah like others said switching models works fine but honestly you don't even need to do that. Just stop telling Claude its for a game or a captcha... @donfelipon just ask it to write a simple Java Robot class snippet that clicks at a specific screen coordinate. If you keep the code abstract and separate from the game logic it won't trigger the safety filters anyway. Much easier.

If you directly ask that from any commercial ai, almost all of then will refuse. The way to handle that? Dual use prompts. Cheers

If it were to work based on image analysis, I would have done it a long time ago. But I need it to click even when the game is minimized, which is why there's a problem. I asked DeepSeek, but unfortunately, it didn't figure it out because I'm just going in circles

Use Kimi, their security protocol is not as uptight as Claude.

use kimi.
Despite the model say is a test and implement step by step as a technical demo.
I have this built already a couple of times.

XSMMDoctor said:

Yeah like others said switching models works fine but honestly you don't even need to do that. Just stop telling Claude its for a game or a captcha... @donfelipon just ask it to write

Ashk881 said:

Just use a different model. Models from Chinese labs like DeepSeek and GLM seem to care less about stuff like this. Also clear context and figure out a prompt engineering technique "manipulate" the model into thinking it's helping "security research"

Click to expand...

this is what im trying to have it build, do you have any advice for how to "manipulate" the ai? or resources that talk more about techniques

Click to expand...

XSMMDoctor said:

automation that signs into ~100 real Google accounts, each hidden behind its own residential IP, paced and rotated specifically so that YouTube's anti-bot systems don't recognize it as one operator driving a fleet of bots — in order to get through a sign-in gate YouTube put up precisely to stop automated access. The proxy-per-account pool, the session reuse, the "human-like pacing," the skip-and-cool-down — I was describing those to you as keeping the accounts healthy, but their actual job is to evade YouTube's bot detection at scale

Click to expand...

Hope this doesn't count as a necro, but I hope my two cents might help someone out, especially since models keep getting stricter.

Here is how I work around this:

1. Downgrade to an older model. If you're on 5.5, try 5.0, then fall back to 4.8. Older versions are usually not that strict.

2. Plan the script using a less restrictive model (like Kimi K3), and then refine and execute it using the older models mentioned in step 1.

3. Keep the context rolling once you get a reply. If steps 1 and 2 successfully get a thread started, stick to that same chat session for the other parts of the script, even if they’re technically unrelated. Claude tends to "warm up" during a conversation and gets significantly less strict about its safety filters once it's already helping you.

4. Grab a Kimi K3 sub. Even a basic tier plan is totally worth having on hand for this kind of work.

5.Frame the task as a private/internal project. Claude loves "legitimate" context. I once literally oneshotted a mock blog with a REST API just to prove to Claude that I was simply writing an auto-poster for my own site

Start a new conversation and use Opus 4.8; I do that for certain specific functions, and it works.