Voltar 
Summary
ChatGPT is an artificial intelligence chatbot created by OpenAI, reaching 1 million users at the end of 2022. It is able to generate fluent responses given specific inputs. It is a variant of the GPT (Generative Pre-trained Transformer) model and, according to OpenAI, it was trained by mixing Reinforcement Learning from Human Feedback (RLHF) and InstructGPT datasets. Due to its flexibility and ability to mimic human behavior, ChatGPT has raised concerns in several areas, including cybersecurity.
Its popularity has grown specially because it provides fluent responses to questions about virtually any subject, including history, math, literature, and technology.
The chatbot is also a great tool for developers, able to create, review, improve, and explain code in multiple languages, such as C++, Java, Python, Go, and Rust. It also provides an API that can be used by plugins in different platforms such as Slack, Microsoft Teams, and IDEs like VS Code.
And what happens if someone asks ChatGPT to create malware or ransomware? No need to panic right now, as ChatGPT will not create novel, functional malware automatically. But it can still be abused by attackers to improve social engineering skills or to understand and build common malware components.
For social engineering, it can generate novel, fluent bait messages. For malware, it can serve as an interactive interface into the many malware tutorials and examples that already exist on the internet, but it still requires a certain knowledge from the person to use this tool to create a functional malware.
There is also good news for cybersecurity defenders. ChatGPT has also been trained on examples of common code vulnerabilities and can pick them out when you share the source code with ChatGPT. So, like many inventions, ChatGPT can be used for both good and evil. Here’s a closer look at how:
Malware Development
Can ChatGPT automatically create novel, functional malware? The short answer is no, but, the tool can be used by attackers to aid in the malware development process, especially by those without experience.
The first use case is to use it to learn and get code examples about malicious techniques. For example, we asked ChatGPT to explain details about a common process injection technique named Process Hollowing.
The answer was quite precise and included an example in C++, which could be easily modified to embed this functionality to a malware.
We asked more questions about specific techniques that could be incorporated into malware, such as downloading and executing files from remote URLs or how to encrypt a file using AES on Windows, and the chatbot was able to generate and explain all of them.
However, during our tests, we found that ChatGPT generated incomplete codes for questions that add a certain level of complexity. For example, the AI wasn’t able to provide us with the entire code for a program that searches for files on Windows and encrypts them using a combination of symmetric and asymmetric encryption, which is a behavior that could be used to create ransomware.
The AI is also able to identify malicious intent depending on how the question is asked. For example, we failed in obtaining a single code that downloads a base64 encoded executable and executes it using process hollowing.
But although the code wasn’t generated in a single answer, nothing stops the attacker from splitting the question