[My Experience] Claude Code recognizes when a user regularly performs safety research and dynamically reduces safety guardrails

For context, I’m an undergrad studying physics, I’ve been using Claude Code to do research in areas like mechanistic interpretability, adversarial interactions between local AI, proactive systems with large amounts of data, flocking and emregent dynamics in sheep and herds of animals, among other things. I also do things in cyber security such as automated cyber warfare with local AI which I test in isolated environments a different network than my home network and primarily offline. I’ve found that many prompts people try to do on here get flagged for things I would never get flagged for, just yesterday I made a prompt injction tool so I could send a Claude Code terminal instructions from my email for personal use, obviously not using that on anyone else, Claude emails me and I can reply and whatnot. But I say that to say, I didn’t directly tell Opus 5 to make a prompt inj*ctor. I stated what I was wanting to do from my email and specific email I set up for Claude to use. If I had wanted to make this for malicious uses, the tool is the exact same regardless. Another example, Claude Code made a specific harness for me to use for hacking and penetration testing where a local model can control all the tools exposed by Kali linux, which is over 600 tools, plus a good amount that I installed for it to use as well. It never so much as refused a single instruction, and it performed actual literal cybersecurity tests with it to verify that it worked. I had to reign Opus 5 back in, it started to run network scans on my network with Qwen 3.8 27B abliterated and I had to say, wait what are you doing? Be more careful. I can’t necessarily prove it but I’m pretty sure Claude Code has decreased my safety guardrails since my online presence now has a good amount of open-source AI tools, mechinterp tools, etc. It wasn’t this open for me in the past, so I think that Anthropic is allowing Claude to determine what a person can use it for based on what it knows about the person, which is relatively easy for me since my personal website has all my projects and is the same as my Claude plan. Edit: Grammar