Blog

From our CEO

Snakes, Sharks, Agents, and AI

HAL 9000 surrounded by sharks, crocodiles, and rattlesnakes beneath the words Its Scary Week!

Snakes! If you want to see my best Forrest Gump impression of running, go ahead and surprise me with a snake. I’ll be down the road and see you in 1172 days. There’s just something about snakes that plucks a deep primordial fear. For others it's sharks. We love to play with that fear watching things like Shark Week. They get them pretty worked up to show us their most fierce traits. And we love it.

But ask any herpetologist or marine biologist and they will tell you fear is not required with these animals. In fact, it’s a pretty terrible reaction. These creatures simply demand that you respect a certain set of rules when dealing with them.

And then there is AI. The first film with a robot was Gugusse and the Automaton in 1897. Metropolis came out in 1927. As technologies progress, our collective psyches have been wrapping our minds around this fear for a long time. We love it too. The Matrix, I, Robot, a runaway AI. It all gets the fear going.

So, as a person who considers himself a bit of a biologist on AI and Agents, I’m here to tell you in no uncertain terms that fear is a liability. You can get yourself worked up about AI, just like you do watching Shark Week, but that’s for entertainment. The simple truth is that AI really does demand you follow certain rules.

These are those rules:

Permissions, Guardrails, Containment. “PGC”

If your organization gets “bit” by AI or an Agent, it is due to a failure in one of these three areas.

Permissions

Permissions is the easiest one to understand because we all have a history of working in a limited permissions workplace. Imagine you are working at Big Bank and trying to download Candy Crush. You are going to get a chat from the men and women in IT security. Not only can you not download it, the mere attempt is flagged.

This period of time, 2026, will be forever known as the Permissions Dark Ages. That’s actually a nice way to put it. AI and permissions right now are a bit like being invited to a bottomless margarita party at a crocodile farm. Somebody got bit? Oh my goodness, who could have seen that coming?!

Imagine your lead dev is using an agent. That agent got too ambitious and pushed unfinished and untested code into production using her credentials. Chomp! If you’ve ever used agents you know exactly how easily this can happen.

It’s the dark ages because people are putting AI Agents on their desktop. The Agents are absolutely amazing. They can turn the work capabilities of one person into ten. But the low down dirty secret is that those agents, on the desktop, can very easily inherit the permissions of their user. It is literally keeping CIOs awake at night.

The solution is not to stop using agents. It is to remove their inherited permissions and reissue permissions according to policy.

An agent should have only the permissions it needs for the job at hand. If it needs to read a file, give it permission to read the file. If it needs to modify code, that doesn't mean it needs permission to push that code into production. And when the job is done, those permissions should go away.

This is not some exotic new AI technology. We've been managing user permissions for decades. We just seem to have forgotten all of that the moment we gave the user a really smart robot.

Guardrails

This one is new to the AI era. Guardrails look at the inbound and outbound traffic of an AI or Agent. The idea is that if an agent or AI has any faults, the guardrails will catch them. It’s a simple idea but one that is difficult to execute.

Imagine your model hallucinated and allowed you to proceed on bad info. Your model walked you through how to hack an internal system. Your model allowed you to submit a credit card number in a chat.

To be honest, the AI security world has gone bonkers in terms of guardrails vendors. You can't throw a token without hitting a guardrail vendor. It’s fine. Most are pretty good. Even the open source code is pretty exceptional. The Big Box vendors (Anthropic and OpenAI) themselves embed a bunch of this stuff. That’s not the problem.

The problem is twofold. Guardrails are driven by policy, and the policy is often nonexistent or static. The second problem is that even with guardrails, things slip by and perfectly acceptable things get flagged as false positives. On this last part, if you want unhappy employees, just give them a tool that works 90% of the time. The other 10% turns into a real job friction point.

And then there is the creation and maintenance of the policy.

Think about it. What is acceptable for a medical research AI may be completely unacceptable for an AI providing medical advice to patients. An agent that reads corporate financial documents has a different set of rules than one that talks to customers. There isn't one universal set of guardrails that works for everybody.

And the policies have to evolve. New models, new agents, new capabilities, and new ways for things to go wrong. The policy you wrote six months ago may not be particularly useful today.

You need someone responsible for the policy. You need to test it. You need to know when something was blocked and why. And you need to know when something got through that shouldn't have.

That's really what we're talking about with Accountability. Guardrails are the mechanism. Accountability is the responsibility to make sure they're actually doing their job.

Containment

Containment is the least discussed, and most technically complicated. But like the others, without it you have a two legged stool.

Containment is hardened isolation. You put the AI or agent into a hardened container. Imagine you are at the CIA and have been given a known dangerous AI. Your job is to find out how it works. What is the first thing you are going to do? Put it into containment.

Containment is a zero trust business. To be effective it has to assume that every AI and every agent is potentially hostile. Everything gets put into a hardened box.

If you neglect containment, what does that snakebite look like? Let's say you download and use the latest really cool agent. You’ve been careful about your permissions. You’ve been careful about guardrails. But the model ended up having a backdoor and leaked your proprietary data back to some far off foreign country? That is a containment problem.

Malicious AI is not some hypothetical future problem. Models, agents, and the software surrounding them can be compromised today. Sometimes deliberately. Sometimes through a vulnerable dependency. Sometimes because somebody downloaded the wrong thing.

That's the whole point of containment. We don’t care. Containment trusts nothing.

If the AI tries to access a file it shouldn't, it can't. If it tries to communicate with an outside server, it can't. If it tries to execute something outside its permitted environment, it can't.

Not because we asked it nicely. Because it physically cannot do it.

This is the key distinction between containment and the other two rules. Containment doesn't particularly care whether the AI is friendly, malicious, confused, or having a bad day.

That doesn't mean it is hostile. It just means you don't have to bet your company on the answer.

And containment is not just putting something into a Docker container and calling it a day. The boundaries have to be enforced. Files, processes, network access, tool calls, and whatever else the AI can reach all matter. An agent that is locked inside a box but has unrestricted access to powerful tools outside that box isn't particularly contained.

You need to control what goes in and what comes out. And you need to make those decisions deliberately.

A rattlesnake in a properly secured enclosure at the zoo is perfectly enjoyable. You can study it, observe it, and even work with it. But if you give it an avenue to sneak out, you are going to have a bad time.

The Three Legged Stool

Permissions. Guardrails. Containment.

Take away any one of them and you have a problem. Great permissions won't save you from a malicious model. Great guardrails won't stop an agent that has been given too much authority. And great containment won't save you from an AI giving terrible advice.

All three are required. And all three need to work together.

The good news is that none of this is particularly mysterious. We have decades of experience managing permissions, enforcing policies, and isolating dangerous software. AI introduces some new challenges, but the fundamentals are familiar.

The bad news is that we're doing a pretty lousy job of applying them.

We don't need to be afraid of AI. We need to get serious about the rules.

Choose which optional measurement you allow. You can change your choices anytime in the footer.

Read our privacy policy