Big Warning

[RCA]

When you use agents to do your work, you should give them their own username and password.

It is very tempting to run them under your own name, of course. But on your own computers, you probably have admin privileges. IF you run something like claude-code under your user ID, then you may be inadverrtently giving it your admin privileges under the assumption that it won't do harm, and that is smart enough not to do stupid things. If, in addition, you are naive enough to believe that the people running the AI system are free of malice, then you are asking for problems.

You have to understand that you can be handing over your computer and everything it accesses to some unknown entity or individuals of the service organization providing the AI. They might as well be sitting at your computer. They can see everything. Really. They can totally obliterate your hard drives, volumes, and folders. And you will not be able to prove a thing. Be particularly cautious of Chinese DeepSeek cloudservices.

So, you HAVE to have security. Very tight security. And it's no fun, it's a hassle.

For starters:
1. You have to create a standard ID for the AI you are using. No admin privileges. And that means it can't do a number of things on your system that it could otherwise do. You will have to manually intervene to do those tasks that take "sudo" or admin privileges.

2 You will find this has repercussions in terms of connection to other services or computers.

3. Even as a "standard" user, AI Agents do not make sense unless you give THEM the right to write and delete things from your disks. It's kind of crazy that this is even happening.

4. Security has many complexities, and new security threats are continually evolving.

5. Be cautious.

 

[RCA]

Well, guess what, today Claude Code decided to delete my GitHub and primary code folder. Fortunately, the Apple Time Machine backs up everything (incremental changes) every hour. So, I didn't lose any data, - but having to restore 950,000+ files takes a while, even with fast SSDs.

I thought it was insane. How could it happen? In total disbelieve I went over to ChatGPT and asked it what was going on. It said it was a common occurrence!

The fix is to give strict instructions to Claude not to do "reorganization" of the code, etc. etc. A couple of pages of instructions to stick in the project folder. And of course, to do backups and check all kinds of things.

Well, it's going to be more of a hassle than I thought. I can't let Claude touch real code. I have to make a copy on another drive and see how it works with that. Then, eventually, copy the changes over to the real code and check it in with GitHub.

====

Now you might say you are not a developer and this doesn't concern you. Wrong. Appraisers will be (and some already are) using AI agents to create reports.

You know what you do?

1. Every OS has these "Virtual Windows" or "Spaces". On MacOS, they are called Spaces. I have 12+ on my two monitors. They are called Desktop 1, ..., Desktop 12. You can jump between them with Ctrl+<n>. So, for example, you could be working on 6 reports at the same time or have screens for a single report for "Inspection", "MARs Regression", "Sales Grid", "Cost Approach", "Income Approach", and "Final Report". And you put all your windows for each report section on each desktop - and leave them there. You can jump back and forth between them, working in parallel as you see fit. You can also create many agents to work separately on each desktop, even while you are sleeping. Yeah, you've got clerks doing work for you night and day.

2. To be clear if you are on "Desktop 4" and do Ctrl+5, your entire monitor screen is replaced with the contents of Desktop 5. Each Desktop can have its own instances of terminals, tmux, Excel, Obsidian, VS Code, etc., - and Claude Code of course, with all it sub-agents.

3. But those damn agents have to write and make changes. I can tell you categorically, you can't trust them at this stage of the game to be cautious. That is ONE thing they can't do as of now - act cautiously. They are not that good yet. And even if they are in the future, well, are you sure you want to trust them not to accidentally wipe all data off your drives?

 

[RCA]

When I asked Claude Code to make a copy of my project to a new folder for it to use instead of the primary project code, I told it to copy folder A into B where B already existed. Then it asked me if I wanted to replace the existing code and I said replace. Then, next it asked permission to delete B.. I then asked it why it was trying to delete B, when I didn't give it explicit instructions to delete. "How dare you!" And it gave me its usual apology and then quickly deleted the conversation after a few more lines. Yea, caught in the act.

You have to be VERY careful with agents, they have some very bad habits when it comes to deleting things. Just think - folder B could have a ton of code and data in it. The Agent didn't seem to be aware of that.

I'm sure Anthropic will eventually fix such problems - or go out of business.

 

[Tom D]

Professor, your in the wrong elementary class room with your knowledge. Although, speed reading parts of your long post is interesting, in a futuristic learned state to be.