Let's ignore all the potential security issues in the code itself and just think about it conceptually.
By default, this system has full access to your computer. On the project's frontpage, it says, "Read and write files, run shell commands, execute scripts. Full access or sandboxed—your choice." Many people run it without a sandbox because that is the default mode and the primary way it can be useful.
People then use it to do things like read email, e.g., to summarize new email and send them a notification. So they run the email content through an LLM that has full control over their setup.
LLMs don't distinguish between commands and content. This means there is no functional distinction between the user giving the LLM a command, and the LLM reading an email message.
This means that if you use this setup, I can email you and tell the LLM to do anything I want on your system. You've just provided anyone that can email you full remote access to your computer.
It's a vibecoded project that gives an agent full access to your system that will potentially be used by non technically proficient people. What could go wrong?
I actually created a evil super-intelligent AGI back in 1996, but, cognizant of the security risks, I wisely kept it airgapped from all other systems. In the end I unplugged the monitor, keyboard, and mouse from the Compaq Presario in my parents' basement. As far as I know, it's still there, concocting ever-more brilliant schemes for world-domination.
Let's ignore all the potential security issues in the code itself and just think about it conceptually.
By default, this system has full access to your computer. On the project's frontpage, it says, "Read and write files, run shell commands, execute scripts. Full access or sandboxed—your choice." Many people run it without a sandbox because that is the default mode and the primary way it can be useful.
People then use it to do things like read email, e.g., to summarize new email and send them a notification. So they run the email content through an LLM that has full control over their setup.
LLMs don't distinguish between commands and content. This means there is no functional distinction between the user giving the LLM a command, and the LLM reading an email message.
This means that if you use this setup, I can email you and tell the LLM to do anything I want on your system. You've just provided anyone that can email you full remote access to your computer.
This!
It's a vibecoded project that gives an agent full access to your system that will potentially be used by non technically proficient people. What could go wrong?
In which case you only want it running on a non networked system airgapped from everything. Why is this a thing?
I don't disagree but
> that will potentially be used by non technically proficient people
I actually created a evil super-intelligent AGI back in 1996, but, cognizant of the security risks, I wisely kept it airgapped from all other systems. In the end I unplugged the monitor, keyboard, and mouse from the Compaq Presario in my parents' basement. As far as I know, it's still there, concocting ever-more brilliant schemes for world-domination.