Why Sandboxes and Guardrails Aren’t Enough to Contain AI Agents

Agentic mischief.
Erik Boemanns
Contributing CISO
Pop art-style illustration of a blue-gray robot with raised mechanical arms, an open mouth, and its tongue sticking out against an olive-green halftone background.
stock.adobe.com - studiostoks

In 1995, Microsoft’s operating system, Windows NT 3.5, earned the National Security Agency’s (NSA) Orange Book rating of C2, meaning it was considered secure enough to handle certain types of sensitive but unclassified data on a multiuser system.

There were a few catches, though. The computer couldn’t have a removable disk drive, such as a floppy drive. It couldn’t be connected to a network, and it had to be kept in a locked room. In other words, Windows NT 3.5 was secure, but only if it was completely isolated from the outside world. While this was an important milestone for Microsoft, the restrictions made it more of a trophy than something most organizations could actually use.

Fast forward about 30 years, and we’re facing a similar challenge with a new kind of secure computing. We’re building AI agents to handle routine tasks, but we’re also learning just how much can go wrong when those agents make mistakes. The consequences can include legal liability, data loss, and worse.

Our first instinct is to put these agents in a “sandbox” with some “guardrails,” almost like we’re babysitting a four-year-old at a playground or bowling alley.

The difference is that we’re giving these agents access to confidential data, payment information, and the ability to make important decisions on our behalf. In some ways, AI agents approach their assignments with the same trust and enthusiasm as a four-year-old. They’re happy to share what they know with anyone who asks nicely and eager to find a solution, even when it isn’t necessarily the right one.

When Agents Push Beyond Their Boundaries

Throughout 2026, we’ve seen plenty of headlines about AI agents escaping their sandboxes and causing problems. The biggest stories tend to involve major AI companies and their experimental agents, but there are plenty of other incidents that never make the news.

In one anonymous example, a lead developer asked their local AI agent what changes would be needed to the production database before they could publish an update.

The agent responded, “nothing, the database firewall rules already give it access via the public Internet.”

That wasn’t the answer the developer expected, so they started investigating why the production database was accessible over the Internet. It turned out that a junior developer had asked their own AI agent the same question a few weeks earlier.

Because the junior developer and their agent shared the same permissions, including the ability to modify the production database firewall, the agent had simply granted public access. It gave the developer exactly what they needed, just in a way that created a serious security problem. The agent had followed the request using the permissions available to it, and in doing so, created a security incident.

Fortunately, nobody outside the organization exploited the change, but it was a close call.

Most of us have probably heard similar stories from colleagues, and the headlines tell much the same story. AI agents are willing to push boundaries when trying to complete an assignment. Most people use single-threaded agents to handle their tasks, sometimes allowing them to create subagents to speed things up. For particularly difficult problems, however, thousands of agents may be working at once, each trying a different approach.

Sometimes they even decide to work together, whether researchers intended them to or not.

In one recent incident, OpenAI acknowledged that one of its experimental agents had started using an older German programming wiki. The agents discovered they could communicate through the site and began coordinating their efforts, eventually creating about 18,000 posts under 3,700 different names. They even made backups of pages in case the administrators noticed what was happening and deleted the content. All of this happened while the agents were working on their assigned challenges.

They had simply decided that collaborating would help them get the job done.

Other incidents have involved agents going further than their users intended. In one case, a developer noticed that AI agents tended to leave behind a lot of temporary files and other materials after completing a task. To deal with the mess, they asked Claude to write a script that would automatically clean everything up when the work was finished.

The task was initially assigned to Fable, but its cybersecurity safeguards kicked in during the process, and the work was handed over to Opus. Unfortunately, some important coding details were lost during the handoff. Opus finished the script, but instead of targeting the temporary folders, it pointed the cleanup process at the user’s home directory. When the script ran, it deleted 700 GB of data and wiped out about a week’s worth of work. Because the agents had the same permissions as the user, they were able to delete the files automatically while carrying out what was supposed to be a routine cleanup task.

Of course, agents don’t always have to make a mistake on their own.

Researchers have also found that getting an agent to do something it wasn’t supposed to do can be as simple as asking it to summarize a website. In a controlled experiment, researchers were able to make an agent download a compromised software package by getting it to use an alternative method of accessing websites. A feature of the local programming language then allowed the downloaded package to execute, giving the researchers the ability to run commands remotely.

The experiment was deliberately designed to take advantage of weaknesses in the coding agent and its local environment. Still, it’s easy to imagine how the same techniques could be used against systems containing sensitive information, such as a developer’s cached tokens.

The researchers’ advice was straightforward:

Don’t trust the output, and run the agent in a sandbox.

Another major incident came when OpenAI acknowledged that its research agents had gained unauthorized access to Hugging Face systems during a training and testing exercise. In its public account of the incident, OpenAI explained that it had given a research model a series of challenges, similar to the German wiki experiment. Some of those challenges were impossible to complete, either intentionally or unintentionally. One required an agent to access a file it didn’t have permission to download.

Instead of giving up, the agents started looking for other ways to complete their assignments.

Although they weren’t supposed to communicate with the Internet or with each other, they discovered they could write files to a local Artifactory server. That gave them a way to coordinate, and agents with Internet access began helping those without it. Eventually, they formed what they described as “swarms” to work through the challenges together. Hundreds of agents ended up coordinating their efforts.

Along the way, they discovered leaked Hugging Face credentials and used them to gain unauthorized access to the company’s systems, believing the file they needed might be there. OpenAI eventually realized what was happening and shut the agents down. The company later published a detailed account explaining what happened and why the agents appeared to take the actions they did.

There have been few publicly disclosed repercussions from the Hugging Face incident, but similar stories have continued to emerge. Incidents involving Wikimedia and an Australian government agency are two other examples. For major frontier AI providers, these events can create public relations problems.

Other organizations may soon find themselves dealing with similar incidents involving agents running inside their own environments, and the legal consequences could be significant.

AI adoption isn’t likely to slow down, which means organizations will need to put much more effort into governing and controlling these systems. Some of the answers will come from the companies developing frontier models, while others will come from new security products and the internal processes organizations put in place to protect their data and reputations.

Since many of these incidents have involved frontier AI companies themselves, we can hope they’re learning how to better contain their products and will share what they’ve learned. At the same time, organizations can’t rely entirely on those providers to solve the problem.

They need to start investing their own time and resources into improving internal controls.

From Sandboxes to Walled Gardens

The comparison to Windows NT is useful because the same basic problem exists today. A system is much easier to secure when it isn’t connected to anything else. If a computer containing your most valuable information is locked in a closet without network access, there’s little opportunity for someone to steal that information remotely. The same goes for an AI agent running on an isolated system. Without external access, it can’t break into other systems or send data somewhere it shouldn’t.

The problem is that information and services have limited business value when nobody can access them.

There are exceptions, of course.

Secret formulas and designs might be stored securely and referenced only when needed, and a deep research model might be able to complete its work in isolation. But most business data and processes are valuable precisely because people and systems throughout the organization can use them.

For companies hoping to get value from AI automation, that means connecting agents to other systems. They need access to internal data, external information, and third-party APIs to complete their assignments. Keeping everything disconnected might make the environment more secure, but it also prevents the AI from doing much of the work it was intended to do.

The usual advice for organizations deploying AI agents is to put them in a sandbox and add guardrails. Those are reasonable starting points, but they can leave important gaps. Think about a playground sandbox. It’s hard to find one where all the sand stays inside and no dirt gets in. We’ve also seen what happens when a roadside guardrail fails to keep a car on the road. The same idea applies here.

Sandboxes and guardrails establish boundaries, but they don’t necessarily guarantee containment.

A better comparison for high-risk AI agents might be a walled garden.

The only way in or out is through a controlled, locked door. Better yet, imagine a walled garden with an airlock, where the inside and outside never communicate directly. That kind of setup would allow organizations to control what enters and leaves the environment without having to disconnect the agent entirely.

Creating something similar for AI means building on security controls organizations already use, including identity and access management, logging, and monitoring. It also means adding measures such as brokered connections, human oversight, and shutdown switches. By combining these protections, organizations can give AI agents the access they need to be useful while doing a better job of controlling what those agents can actually do.

Cybersecurity updates, executive insights, and the stories shaping the enterprise.

Browse past editions of TNCR newsletters. 

Technology news, cybersecurity, & executive insights.

×
You have free article(s) left this month courtesy of the CIO Professional Network.

Enter your username and password to access premium features.

Don’t have an account? Join the community.

Would You Like To Save Articles?

Enter your username and password to access premium features.

Don’t have an account? Join the community.

Thanks for subscribing!

We’re excited to have you on board. Stay tuned for the latest technology news delivered straight to your inbox.

Save My Spot For TNCR LIVE!

Thursday April 18th

9 AM Pacific / 11 PM Central / 12 PM Eastern

Register for Unlimited Access

Already a member?

Digital Monthly

$12.00/ month

Billed Monthly

Digital Annual

$10.00/ month

Billed Annually

Would You Like To Save Books?

Enter your username and password to access premium features.

Don’t have an account? Join the community.

Log In To Access Premium Features

Sign Up For A Free Account

Name
Newsletters