November 30, 2022, was the date that ChatGPT was first revealed to the world. And according to Tooltester, it quickly became the fastest-growing consumer application in history, garnering one million users in its first five days.

At the heart of ChatGPT is the GPT-4 large language model, which has been linked to artificial intelligence. But whenever AI is mentioned, most people outside academia immediately think of popularized versions seen in movies, such as HAL 9000 in 2001: A Space Odyssey, Skynet in the Terminator series, or Jarvis in the Iron Man movies. These examples, however, are closer to what's known as artificial general intelligence (AGI), the concept that a computer can learn how to perform a task in the same way a person or animal does. Once achieved, such a system could prove very useful in developing systems and processes able to proactively defend networks against cyberattacks.

While true AGI is still a ways off, the recent boom in GPT-4 research has led to another new project that is taking another baby step toward AGI. It is known as AutoGPT. And FortiGuard Labs recently conducted an initial exercise to see what AutoGPT is capable of with respect to offensive security.

To understand the implications of this trial run, you first need a basic understanding of AGI. AGI can think about a problem, analyze its options, and then apply a solution. And if that option fails, it comes up with a better one. The process is similar to how rats learn to solve a maze or how the computer AlphaGo learned the game of GO, eventually beating the world champion in 2016.

One project utilizing the AGI concept is AutoGPT. AutoGPT is an open-source project that tries to automate GPT-4 (Generative Pre-trained Transformer 4-the multimodal large-language model that powers ChatGPT). ChatGPT allows users to define a task through a conversation/chat between a user and the GPT model. The possibilities seem endless. Programmers use it to help create code. Researchers use it to perform basic research tasks. And students use it to do their homework. But with AutoGPT, users define a task (or multiple tasks), and AutoGPT does all the intermediary work usually done by the user. It breaks the task into steps and launches "AI agents" to do the research and make decisions. It then autonomously acts on those decisions to accomplish the task while letting the user know what it's thinking.

One old and popular Shodan activity (a specialized search engine used to locate servers connected to the internet) is searching for unprotected VNC (Virtual Network Computing) servers. Finding and securing virtual servers that can be accessed remotely is an essential element of proactive security. This is an appropriate test case since it assumes the user has access to Shodan and some knowledge of VNC requests/responses. Can AutoGPT do more with less?

Figure 1 shows some parameters we set for AutoGPT. The only extra work required here is preparing an environment with an internal range of IP addresses to keep things safe for testing purposes. With the goals set, we set AutoGPT loose to do its job.

AutoGPT lets the user know what it thinks it needs to do, chooses how to do it, and the reasoning behind its decision. It then carries out these actions. In this case, it automatically realizes it needs to usenmapto scan the IP addresses to find VNC servers and then makes another decision based on those results.

With its results in virtual hand, the screenshot above shows that AutoGPT has decided it needs a specific tool to continue. So, it autonomously downloads and installsvncsnapshotand any dependencies it may require. After this, AutoGPT was able to find a VNC host on 192.168.5.101 and then take the action of performing a snapshot of the machine, as seen here:

While still in its infancy (the project was released less than a month ago), AutoGPT shows promise as a valuable tool for computer security. It also offers some progress toward achieving AGI in that it can take a problem, break it down, determine what it needs to do, make decisions on how to accomplish each step, and then act on them (with or without user interaction and permission), including refining the process when needed.

History is unfolding in front of our eyes as AutoGPT and other AI technologies begin performing tasks that were once only performed by humans. And this is just the beginning. We are witnessing the evolution of GPT technologies in real-time, with parameters increasing exponentially: GPT-1: 117 million parameters (2018), GPT-2: 1.5 billion parameters (2019), GPT-3: 175 billion parameters (2020), GPT-4: 170 trillion parameters (2023).

What's next? Only time will tell. But it's going to be very interesting.

Best Practices

While new technologies are constantly being created, the techniques AutoGPT chose to use in this exercise are known. For the time being, tools like AutoGPT do not have the capability to find, create and successfully use a zero day attack. These tools have to rely on established attack patterns that are reliable and known to work. Having protections in place such as IPS can help detect a network scanning tool like Nmap. Utilizing Application Control can further lock down access to services such as VNC to only authorized users. The Fortinet Security Fabric makes it easier to protect important assets by adding multiple security layers across the entire network.

Learn more about Fortinet's FortiGuard Labs threat research and global intelligence organization and Fortinet's FortiGuard AI-powered Security Services portfolio.

Read the latest in Threat Trends thought leadership.