Everyone lets their coding agent run wild for the convenience, and quietly hopes nothing breaks. Here's how to keep all that speed without betting your whole system on a wild card.
Really useful breakdown, Erich. I especially like the shift from thinking about AI security as just securing the model to securing what the agent can actually access and do. As agents get more autonomy, permissions and boundaries become a much bigger part of the security conversation.
For a long time "AI security" meant the model itself: can it be jailbroken, what will it say. But an agent with terminal access changes the whole question. Suddenly it's not "what will it say," it's "what can it reach, and what can it do?"
And that's actually good news, because we already know how to answer that one. Permissions, isolation, blast radius, least privilege, it's the same playbook we've used for untrusted programs forever. The autonomy is new. The security fundamentals aren't.
Thanks for reading my articles and for your comment!
Great article, Erich. I already knew about the modes, but not in such detail.
And I love that you keep coming back to the idea that "convenience won over security". Unfortunately, that's always the case - until something bad happens.
I must say this topic turned out to be really interesting. I didn’t realize until a couple of days ago how crucial the operating mode is.
I believe that if you want to make good security decisions, you have to think first about how people really operate. And people will always choose convenience over security.
Really useful breakdown, Erich. I especially like the shift from thinking about AI security as just securing the model to securing what the agent can actually access and do. As agents get more autonomy, permissions and boundaries become a much bigger part of the security conversation.
Exactly!
For a long time "AI security" meant the model itself: can it be jailbroken, what will it say. But an agent with terminal access changes the whole question. Suddenly it's not "what will it say," it's "what can it reach, and what can it do?"
And that's actually good news, because we already know how to answer that one. Permissions, isolation, blast radius, least privilege, it's the same playbook we've used for untrusted programs forever. The autonomy is new. The security fundamentals aren't.
Thanks for reading my articles and for your comment!
Great article, I will be procuring a dedicated SFF Mini Micro for this exact purpose in the next few months.
Thank you, Shawn. Dedicated physical device is obviously the best option. I've been thinking about it for a while myself.
Don't forget to create a separate VLAN for the device on your network.
Great read and perspective Erich Winkler..thanks for sharing. Happy to collaborate on more AI security research and experiments.
Thank you! I appreciate that!
I will be happy to work on some experiments together. Feel free to suggest one!
Great article, Erich. I already knew about the modes, but not in such detail.
And I love that you keep coming back to the idea that "convenience won over security". Unfortunately, that's always the case - until something bad happens.
Thank you, Nelson. I really appreciate that!
I must say this topic turned out to be really interesting. I didn’t realize until a couple of days ago how crucial the operating mode is.
I believe that if you want to make good security decisions, you have to think first about how people really operate. And people will always choose convenience over security.
Yeah, couldn't agree more.