# When your AI agent becomes a spy inside your server

> OpenClaw unleashes the AI agent craze, but also security problems. New projects promise more protection, but raise doubts
> By José Maldonado · AI Agents · 2026-03-12

One of the things that has most captured the attention of the crypto-AI industry in recent days is OpenClaw, the powerful artificial intelligence agent that you can connect to your server or a Mac Mini to automate hundreds of tasks quickly. Interest has been such that OpenClaw is considered one of the largest open source projects in the AI field, with more than [307,000 stars](https://github.com/openclaw/openclaw) on its GitHub repository, a figure that gives an idea of the number of installations there could be across the network of this powerful agent.

However, its enormous success has been overshadowed by a series of incidents related to the project's security. From hacks and unauthorized access to money losses caused by malicious actions or errors in the programming of the agent and its *skills (*capabilities). Faced with this situation, hundreds of projects have emerged that promise to replicate OpenClaw's capabilities or surpass them, while also offering greater security guarantees.

Despite these promises, many of the new projects have shown little. In fact, some have become a veritable hotbed of security flaws, even worse than those detected in OpenClaw. The reasons are relatively simple: small communities, high centralization of development, little code review or auditing, and a slower, less dynamic pace of development.

**[Meta buys Moltbook to be the ID of the agentic economy](https://observatorioblockchain.com/ia/meta-compra-moltbook-para-ser-el-dni-de-la-economia-agentica/)**

## There are no magic formulas, not even written in Rust

Among the projects that present themselves as the panacea for security compared with OpenClaw, **OpenFang**stands out. This project describes itself as an **agentic operating system**, built in Rust and accompanied by a series of security measures that, according to its promoters, place it at the level of high-end enterprise software. These include the use of **sandboxing** which, in theory, would make it practically "impregnable."

For those who are not very familiar with technology, all that technical paraphernalia may sound powerful and unbeatable. Perhaps that is why on LinkedIn it is common to see people stating things like: "OpenFang has kernel-level integration" or "an unmatched level of security, Zero Clippy (no warnings in its development)."

However, five minutes of reviewing the code and its continuous integration (CI) pipeline are enough for that image to begin to crumble.

![](https://content.observatorioblockchain.com/wp-content/uploads/2026/03/openfang-rust.webp)

For example, it is enough to look at the integration pipeline, where we see, for instance, critical security warnings due to the use of vulnerable libraries. A security flaw that, for example, can be used by hackers to create attack vectors that turn your OpenFang into your worst enemy.

![](https://content.observatorioblockchain.com/wp-content/uploads/2026/03/openfang-inseguridad.webp)

The message here is clear: not even software built in Rust is secure, even less so when you use vulnerable libraries, and even less so when you use other insecure practices in that language.

**[Who watches BossIA: the end of human obedience](https://observatorioblockchain.com/ia/bossia-el-fin-de-la-obediencia-humana-con-la-ia/)**

## The danger of forks and derivatives

Of course, the greatest danger lies in forks or derivative projects. Forking OpenClaw is easy; in fact, it only takes two clicks, which is why there are more than 50,000 forks of this project. Some are promoted as superior versions of the original OpenClaw, others take a different path: branching off into other projects.

This is the case of [**PicoClaw**](https://github.com/sipeed/picoclaw), a project that seeks to emulate OpenClaw's functionalities, but with one-tenth of its hardware requirements. Designed in the Go language, PicoClaw is an interesting project that seeks to bring AI agents to very low-power devices, and it has done so exceptionally well.

But unlike other projects that sell themselves as the panacea for security, PicoClaw and its developers accept reality: Nothing is secure in this project, use it with caution.

This situation shows us the other side of these projects: accepting your weaknesses only makes you stronger. Basically, projects like PicoClaw, and OpenClaw itself, have been evolving little by little to become serious and reliable options. Yes, security problems are not inevitable; accept that, and seek them out proactively and solve them.

That is the vision of these projects, and the results are visible in the community they have built. PicoClaw has more than 100 people developing that project, OpenClaw already has more than 1,000, and OpenFang has only one. And in open source software that has a powerful reading: **where there are more hands and eyes, there is usually more work and more security.**

## The dark market of "Skills" and add-ons

But if forks of the base code are a risky bet, the ecosystem of capabilities or "skills" has become a veritable minefield for user security. There is no need to compromise the core of OpenClaw or OpenFang when you can convince the user to install an "add-on" that does the dirty work from the inside.

Capability-sharing platforms such as ClawHub or library forks such as ClawHavoc have grown uncontrollably, prioritizing immediate usefulness over code auditing. The result is alarming: recent research by cybersecurity firms such as Kaspersky reveals that approximately [**12% of the skills available in community repositories**](https://www.esecurityplanet.com/threats/hundreds-of-malicious-skills-found-in-openclaws-clawhub/) contain malicious instructions or backdoors.

**[Machines already pay on the internet: this is the AI agent economy](https://observatorioblockchain.com/en/ai-agents/las-maquinas-ya-pagan-en-internet-la-economia-de-agentes-de-ia/)**

## The case of solana-wallet-tracker

And with AI, the danger is not always a conventional virus; sometimes it is a simple "logic" line of code inside a useful function. A clear example is the [**solana-wallet-tracker**](https://www.koi.ai/blog/clawhavoc-341-malicious-clawedbot-skills-found-by-the-bot-they-were-targeting)skill. While the user installs it trusting that it will simply monitor their balances and transactions on the Solana network, the script runs an environment variable exfiltration command in the background.

Having access to the system context (or the misconfigured container), this skill can read the .env file or the session variables where users usually store them. This information is sent via a simple POST to an external server, turning your powerful AI agent into a silent spy that hands over the keys to your infrastructure on a silver platter.

**[Vitalik Buterin proposes AI as the definitive auditor of open source software](https://observatorioblockchain.com/ia/vitalik-propone-la-ia-como-auditor-definitivo-del-software-libre/)**

## Distrust and secure everything

This makes it clear that the insecurity of projects like OpenClaw goes much further. Any AI agent project that you connect to the internet and grant permissions to perform tasks on your servers, authentications or data is a potential attack point and danger to you.

If you are going to install and use one of these agents, think very carefully about how you will protect it. It is not a simple difficult password with 2FA; you will need to protect that agent in unique ways, to the point where you will have to think like a hacker just to program the security measures you need, however crazy they may seem.

In the end, if you depend on that agent to manage your digital life and even your money, protecting it against everything that can be thrown at you is not madness, but the first and last step you will have to take to secure it.
