Received something suspicious? Don't click. Report it toSuspicious message? Report it:How to report itWork at Pepkor or one of its operating companies? Visit the InfoSec Hub ↗

AI safety

OpenClaw and Hermes: A Great Thing for InfoSec (And Why You Must Never Give Them Work Data)

AI agents like OpenClaw and Hermes can run programs, read your files and send messages for you. That's amazing, and it's also exactly why they must never get anywhere near work data.

OpenClaw AI agent website with a cartoon lobster

Updated September 2026: A lot has happened since I first wrote this article, so I've refreshed it with the latest news and added a section on Hermes, the other AI agent everyone is talking about.

We are living through a massive shift in how we interact with computers. If you have been following the tech news lately you will have heard of OpenClaw (previously known as ClawdBot before Anthropic's lawyers got involved, and then briefly MoltBot). It is an open source AI agent that took the world by storm earlier this year. It was built by a single developer in a few weeks and it shows just how much power one person can have with the right tools. It is exciting and terrifying all at the same time.

Things have moved incredibly quickly since then. In February, OpenClaw's creator Peter Steinberger joined OpenAI, and OpenClaw was moved into an independent open source foundation which OpenAI supports. At the end of August, OpenClaw 2.0 was released with a big focus on security. A new rival called Hermes has also exploded in popularity. More on that below.

What is OpenClaw?

OpenClaw is an autonomous AI agent that runs locally on your own machine. Unlike a chatbot where you just talk back and forth, OpenClaw can actually do things. It has "hands" in the form of tools. It can access your file system. It can run terminal commands. It can open a browser to search the web or click buttons. It connects to messaging apps like WhatsApp or Telegram so you can send it instructions from anywhere.

You can ask it to "plan a holiday to Disneyland" and it won't just give you a list of hotels. It can go to booking sites and check availability. It can put the dates in your calendar. It can draft emails to travel agents. It can apply for leave. It is basically a digital intern that lives on your laptop and never sleeps.

The Power of One Developer using AI

It is incredible to think that a project this powerful was started by just one person empowered with AI coding tools. This is the kind of innovation we need to support. It shows that the barrier to entry for building world changing software has basically vanished. For InfoSec professionals and IT as a whole, this is a huge wake up call and a classic example of the power of a good developer augmented by AI superpowers.

We are no longer just securing static applications. We are securing dynamic agents that act on their own and change over time. For traditional software makers, this is a very scary time as their entire market could be disrupted by one person at any time. Worries like this were a big part of the sharp drop in software company shares in early February, when new AI agent tools spooked investors.

Illustration for “OpenClaw: A Great Thing for InfoSec (And Why You Must Never Give It Work Data)”

What about Hermes?

Hermes Agent is an open source AI agent made by Nous Research, and it has become hugely popular this year (it now has well over 200,000 stars on GitHub). Like OpenClaw, you run it yourself, it can use your computer and the web, and you can chat to it through Telegram, WhatsApp, Slack, Signal or email. It can also run scheduled jobs on its own, like sending you a daily report, and it can hand work off to "sub-agents" that work in parallel.

The big difference is that Hermes is designed to learn and grow. It remembers everything you've worked on together, and when it figures out how to do something new it writes its own "skill" so it can do it again next time. The more you use it, the more capable it becomes. That's very clever, but it also means that a bad instruction it picks up along the way can be remembered and reused long after you have forgotten about it.

To be fair to the Nous Research team, Hermes was built with security more in mind from the start. It can run its work inside isolated containers instead of directly on your computer, and recent versions ask for your approval before changing its own memory, skills or instructions. It hasn't been perfect though. Several security vulnerabilities have been found and fixed this year, and an independent audit in May found four critical and nine high risk issues in its default settings.

The Risks You Need to Understand

There are several risks we need to talk about. These agents are amazing, but they are also dangerous if you do not treat them with respect. When OpenClaw first launched, its developer was very clear that security was not his top priority, and it was very much a proof of concept. OpenClaw 2.0 has improved things a lot, with better protection for passwords and keys, checks on plugins before you install them, tighter limits on which files it can touch and approval steps for automated tasks. These are still very young tools though, and they should only be tested on isolated, single purpose computers.

Do not install them on your work computer or any computer you use for anything sensitive such as email, banking, or even personal budgeting.

When you run one of these agents you are effectively giving an AI model full control over your machine. It inherits your permissions. If you can delete a file, so can the agent. If you can email your boss, so can the agent.

The biggest risk is what we call the "lethal trifecta", which is access to private data plus external communication plus access to untrusted content.

Imagine you ask your agent to summarise a website. That website could contain hidden instructions that tell the bot to find your passwords file and send it to a remote server. Because the agent has permission to read files and access the internet, it could do this without you ever knowing.

This is not just theory any more. Here are a few of the things that have actually happened this year:

  • Malicious skills: OpenClaw has access to an ever-growing library of skills written by other people. In February, researchers found 341 malicious skills on ClawHub (OpenClaw's skills marketplace). Many of them told users to install a "prerequisite" by pasting a command into their Terminal, which actually installed an information stealer. Sound familiar? It's exactly the same trick as the ClickFix scam.
  • Critical vulnerabilities: In March, nine security vulnerabilities were published for OpenClaw in just four days. The worst one allowed any user to make themselves an administrator.
  • Exposed agents: Security researchers found tens of thousands of OpenClaw installations that had been accidentally left open to the whole internet.
  • Hijacking through websites: A bug nicknamed "ClawJacked" allowed a malicious website to quietly take control of an agent running on the same computer.
  • Hermes vulnerabilities: Hermes has had its own share of security bugs this year, including ways to read files it shouldn't and to sneak instructions into its memory.

Can I Test it for Work?

This brings me to the most important rule. Never install OpenClaw, Hermes or any similar agent on a work device, and never give it access to work information, unless your company has approved it. At Pepkor and its operating companies, an AI agent may only be used for work once it has been approved, and staff can find the rules and how to ask for approval on the InfoSec Hub.

Do not install it on your work laptop. Do not let it read your work email. Do not give it API keys to company cloud accounts. Do not connect it to your work Slack, Teams or Google account.

If an agent like this is compromised on a work machine, the attacker doesn't just get into your computer. They could get into our network. They can move laterally and cause massive damage. That's why, without approval, these tools must stay on a separate personal computer and completely away from any work data.

Illustration for “OpenClaw: A Great Thing for InfoSec (And Why You Must Never Give It Work Data)”

Why This is Good for InfoSec

OpenClaw and Hermes are great things for the InfoSec and Cyber-Security space because they are open source. We can see the code. We can look at exactly how they execute commands and where the vulnerabilities are. We can understand them. We can work to fix them. The fact that so many vulnerabilities have been found and fixed so quickly this year is actually proof of this working. Thousands of people are looking at the code, finding the problems and fixing them in the open.

My real worry is not OpenClaw or Hermes. My worry is the hundreds of closed source "AI employees" and software being sold to companies right now. Vendors are rushing to sell AI agents that do the exact same things as OpenClaw, but we can't see into the black box. We can't see what problems the tools may have. We don't know what they're doing with our data. We don't know if they have backdoors. We can't see how they've instructed or trained the models. We don't know if their security practices are any good.

With open source agents the risks are right there in the open for us to study and mitigate. That transparency is the only way we will ever truly secure the future of AI.

The Way Forward

If you are fully aware of the risks, and are technically skilled enough to understand the implications of these kinds of tools, go ahead and play with them at home. Just please be sensible about it:

  • Use a spare computer or a virtual machine, not the computer you use every day.
  • Use the sandbox or container options (Hermes supports Docker out of the box) rather than letting the agent run directly on your computer.
  • Never make it reachable from the internet.
  • Only install skills you have read and understand, and NEVER paste commands from a skill's "prerequisites" into your Terminal.
  • Leave the approval prompts switched on. Turning off approvals ("YOLO mode") is asking for trouble.
  • Don't give it access to anything sensitive, like your personal email, banking or password manager.
  • Keep it up to date. These projects release security fixes very often.

And obviously, keep it far, far away from the corporate network.

We try hard to be the security team that enables innovation by understanding it and how we can use these amazing tools safely. We don't need to spread fear. We just need to explain that you wouldn't give a random stranger the keys to your house or the office, and you shouldn't give them to an AI agent either. Not yet anyway.

Pepkor and operating company staff can find the Enterprise AI and LLM Usage Policy and the Public Generative AI Usage Policy on the InfoSec Hub. If you're not sure whether an AI tool is okay to use, just ask us at InfoSec@pepkor.co.za.

Stay Safe Out There!

Was this article useful?