This is the piece I'll point people to when they ask where to start — the "vocabulary changed, defensive thinking didn't" line is exactly right. The one place I'd push further: there's a single genuinely new primitive under all of it, and it's that the agent can't distinguish data from instructions. Every classic control still applies, but that one property is why "just validate the input" doesn't fully save you — the malicious input IS valid data that happens to also be read as a command. So least privilege and blast-radius thinking stop being best practice and become the whole ballgame, because you can't prevent the confusion, only contain what it's able to do. Your four closing questions are the actual interview. Great collab
In my opinion, the fundamentals we have known for decades still serve us really well. The only difference is that executing them now has to be far more precise. A loose permission or a missing log used to be a gap.
With an agent that can act on its own, it is an open door.
Looking for a Cybersecurity writer job...if you have any recommendations, please let me know. I have 13 years old total writing experience and around 1.5 years of Cybersecurity writing experience. I am also pursuing a course in Cybersecurity.
Three takeaways that are haunting me: we mistake naming a threat for neutralizing it, we obsess over triggers instead of systemic environments, and true security means making failure boring.
Excellent piece.
Cigarettes ain’t enough to calm the nerves after this one haha.
All of those are true, Phil. In fact, the true security is often very boring. Most attacks can be prevented by very simply rules and cybersecurity fundamentals that people just don't like to follow.
"The architecture should make that failure boring" is the line I would hand to anyone rolling out agents. It moves the question from whether the model can be fooled to what it was allowed to do once it was, and that second question is one security teams already know how to answer.
Exactly. It often surprises me how people approach AI security like it’s something brand new. Yes, there are some specific aspects to that, but the fundamentals stay the same!
The point about looking beyond the vulnerability name and asking about permissions, trust boundaries, data access, and blast radius was especially useful. Appreciate both ToxSec and Erich for making AI security practical and grounded in security fundamentals!
Thank you for your kind comment. We appreciate it.
I’ve said this already, but it often surprises me how people approach AI security as if it were a completely new problem. I know there are some specific aspects to it, but the fundamentals still apply.
As many SMBs and startups are facing a dire need to move from POCs to actual stable (and valuable) production solutions , before addressing the security issues that Erich and Chris highlighted here, they first need to be acutely aware of those.
Thanks for your efforts to educate and evangelize these important topics!
That POC-to-production jump is exactly where it bites: the prototype ships with the same wide-open permissions it had on day one, and nobody pauses to ask what it can now reach in a real environment.
Really appreciate you reading it and the kind words!
It’s precisely the way I built Exec Room, with granular settings per room for your agents. And warnings if anything touches outside or settings are too broad.
I am happy to hear that there are people out there who follow the best practices and create systems that think about security in the design phase. Not as something that will be added "later". (never)
This is the piece I'll point people to when they ask where to start — the "vocabulary changed, defensive thinking didn't" line is exactly right. The one place I'd push further: there's a single genuinely new primitive under all of it, and it's that the agent can't distinguish data from instructions. Every classic control still applies, but that one property is why "just validate the input" doesn't fully save you — the malicious input IS valid data that happens to also be read as a command. So least privilege and blast-radius thinking stop being best practice and become the whole ballgame, because you can't prevent the confusion, only contain what it's able to do. Your four closing questions are the actual interview. Great collab
Thank you, that is a great point.
In my opinion, the fundamentals we have known for decades still serve us really well. The only difference is that executing them now has to be far more precise. A loose permission or a missing log used to be a gap.
With an agent that can act on its own, it is an open door.
Looking for a Cybersecurity writer job...if you have any recommendations, please let me know. I have 13 years old total writing experience and around 1.5 years of Cybersecurity writing experience. I am also pursuing a course in Cybersecurity.
Hello Preethi, thank you for your comment! Unfortunately, I don't know about writer job offers.
However, if you want to learn more about cybersecurity, you can find plenty of resources on Decoded Security!
You can also reach out to me anytime!
Three takeaways that are haunting me: we mistake naming a threat for neutralizing it, we obsess over triggers instead of systemic environments, and true security means making failure boring.
Excellent piece.
Cigarettes ain’t enough to calm the nerves after this one haha.
All of those are true, Phil. In fact, the true security is often very boring. Most attacks can be prevented by very simply rules and cybersecurity fundamentals that people just don't like to follow.
Thank you! We appreciate that!
"The architecture should make that failure boring" is the line I would hand to anyone rolling out agents. It moves the question from whether the model can be fooled to what it was allowed to do once it was, and that second question is one security teams already know how to answer.
Exactly. It often surprises me how people approach AI security like it’s something brand new. Yes, there are some specific aspects to that, but the fundamentals stay the same!
2 of my favorite creators, loved the post guys!
Happy to hear that!
The point about looking beyond the vulnerability name and asking about permissions, trust boundaries, data access, and blast radius was especially useful. Appreciate both ToxSec and Erich for making AI security practical and grounded in security fundamentals!
Thank you for your kind comment. We appreciate it.
I’ve said this already, but it often surprises me how people approach AI security as if it were a completely new problem. I know there are some specific aspects to it, but the fundamentals still apply.
As many SMBs and startups are facing a dire need to move from POCs to actual stable (and valuable) production solutions , before addressing the security issues that Erich and Chris highlighted here, they first need to be acutely aware of those.
Thanks for your efforts to educate and evangelize these important topics!
Well put, Andrei.
That POC-to-production jump is exactly where it bites: the prototype ships with the same wide-open permissions it had on day one, and nobody pauses to ask what it can now reach in a real environment.
Really appreciate you reading it and the kind words!
Spot on, position your security around what could happen (cause the likelyness something will increases with every new model).
Exactly. You can keep lowering the likelihood, but you will never get it to zero. And as you said, the probability seems to be increasing lately.
So the real leverage is on the impact side: assume it gets fooled, then make sure a fooled agent cannot reach more than absolutely necessary.
It’s precisely the way I built Exec Room, with granular settings per room for your agents. And warnings if anything touches outside or settings are too broad.
I am happy to hear that there are people out there who follow the best practices and create systems that think about security in the design phase. Not as something that will be added "later". (never)
No injection in my case, the agent was just confused.
One audit of my Claude Code logs showed my deny hook blocking 40 dangerous commands. 14 of them were rm -rf on home or root.
So the limit sits in a hook now, not in the prompt.