At this point it’s pretty easy to sandbox your agent so it can’t do this. Even a trivial docker container with a shell script would have prevented this.
While i understand you would think a vendor like Google would do sandboxing by default, the industry standard is that your agent gets your user permissions in full. Your external drive doesn’t need escalated permissions to wipe, and the AI companies don’t care at all about your security.
Sandboxing wouldn’t have helped as he asked to create an executable that deletes files. It’s not the agent that deleted the files but the “uninstall” routine in the compiled program that deleted everything.
With such a generic prompt I wonder what kind of mess it has generated.
Probably such a program didn’t need an “install” or “uninstall” routine, as simple programs are portable, don’t need to leave libraries on the disk. He even didn’t know that uninstall and install are the last thing you do in a program, and It’s not integrated in the code, but it’s packaged after the build with third party stuff like nsis or innosetup
At this point it’s pretty easy to sandbox your agent so it can’t do this. Even a trivial docker container with a shell script would have prevented this.
While i understand you would think a vendor like Google would do sandboxing by default, the industry standard is that your agent gets your user permissions in full. Your external drive doesn’t need escalated permissions to wipe, and the AI companies don’t care at all about your security.
The T in AI stands for trust.
Sandboxing wouldn’t have helped as he asked to create an executable that deletes files. It’s not the agent that deleted the files but the “uninstall” routine in the compiled program that deleted everything.
With such a generic prompt I wonder what kind of mess it has generated.
Probably such a program didn’t need an “install” or “uninstall” routine, as simple programs are portable, don’t need to leave libraries on the disk. He even didn’t know that uninstall and install are the last thing you do in a program, and It’s not integrated in the code, but it’s packaged after the build with third party stuff like nsis or innosetup