Apparently my AI agents look like hackers now. 😂
I tell Codex or another agent what I want.
Then it starts opening shells, running scripts, installing packages, reading files, starting processes, using Git, running tests, sometimes spawning other tools.
For me this is just normal development now. I asked for something. They do what they need to do to get there.
But if you look at all of this from the outside, without the context, it really does look like someone broke into the machine. 😂
I read an article from The Hacker News and immediately thought about my own setup with Codex CLI, Herdr, multiple agents, worktrees and automations. And I think this is going to become a pretty big problem.
Because before, security systems could ask:
Did the user execute this?
Now the answer can be:
No. The user told an AI agent what he wanted, and the agent decided to execute this.
That changes quite a lot.
Personally I don't want less autonomy. Quite the opposite. I want agents to do more and more by themselves.
But then we need to get much better at controlling what they can access, isolating them, managing credentials and understanding why they are doing something.
Because apparently the future of software development is me managing a small army of processes that occasionally looks like a cyberattack.
Seems fine. 😂