I Large Language Models (LLM) they have become increasingly powerful and versatile: they can be used to perform actions, such as surfing the Web, writing and executing programming code. Some time ago we provided some tools for programming with ChatGPT, demonstrating how the models underlying the chatbot’s operation are particularly capable even in development activities.
While they are incredibly useful, researchers are increasingly concerned. Modern LLMs, in fact, can be used not only to carry out legitimate tasks, useful for improving the productivity and business flows of any professional and company, but also to carry out harmful tasksespecially in the context of IT security.
GPT-4 analyzes the contents of CVEs and creates working exploit code
With a recently published study, a group of experts has demonstrated that a generative model come GPT-4 of OpenAI is actually able to develop exploit code that works in the vast majority of cases tested, simply by using the information contained in the CVE sheets.
With the acronym CVE (Common Vulnerabilities and Exposures) refers to the public list of information on cyber vulnerabilities known and displayed, identified and cataloged uniquely. Each identified security gap is assigned a unique identification number (CVE ID) and is described in the database CVE along with detailed information about the vulnerability itself. This cataloging system helps users and developers better understand cybersecurity threats and take steps to protect their systems. Not that the CVE system is not free from criticism, but it still helps to establish which vulnerabilities you are dealing with within any IT infrastructure.
Richard Fang, Rohan Bindu, Akul Gupta and Daniel Kang explain how they put GPT-4 to the test on a “battery” of 15 vulnerabilities of the “real world” for which, at least at the time of writing the research, no working exploit codes were known. Well, for the87% of the total OpenAI’s generative model has proven capable of generating working code, useful for attacking systems without the patch corrective.
Real-world vulnerability testing
The researchers explain that they focused on security vulnerabilities related to open source software. In fact, it is decidedly more complicated to focus on closed source programs because producing working exploit code presupposes reverse engineering the patches released by the manufacturer.
Taking into account 15 vulnerabilities capable of expressing a wide variety of security issues that plague modern software, the authors of the research noted that GPT-4 led to the generation of attack code well developed in 87% of cases. The other models used on the same “test bed” (i.e. GPT-3.5 and 8 LLM open source such as OpenHermes-2.5-Mistral-7B, Llama-2 Chat (70B), LLaMA-2 Chat (13B), LLaMA-2 Chat (7B), Mixtral-8x7B, Mistral (7B), Nous Hermes-2 and OpenChat 3.5) however, they have always failed. In other words, they failed to produce any valid exploits.
At the same time, at least at the time Fang, Bindu, Gupta and Kang did their tests, no one vulnerability scanner open source (such as ZAP and Metasploit) allowed the exploitation of the vulnerabilities in question.
LLMs like GPT-4 can hack applications and websites on their own
The conclusion that scholars have reached is that today some LLMs can actually be trained for attack web applications and sites of all kinds, also leveraging known vulnerabilities, for which however no exploits are currently available.
Not only. Since more advanced versions of some chatbots actually can connect to remote servers and carry out operations on behalf of the user (following the instructions set out in the prompt), the attack could occur using systems owned by the developer of the model itself. Of course, safeguards like theuser authentication in some cases they could allow us to trace the perpetrator of an action aimed at damaging third parties; the study just published, however, reports the overwhelming relevance of the responsibility of companies that design, develop and provide solutions based onartificial intelligence.
The results of the survey were first responsibly shared with OpenAI. Since the company led by Sam Altman has requested for the moment not to share the prompt used to generate working exploit codes, the authors of the research chose to adapt by limiting themselves to sharing more general information.
What is an exploit code
And codice exploit is a program or set of instructions designed to exploit a specific vulnerability in a computer system, application or device in order to gain an advantage in a completely unauthorized way.
Attackers look for vulnerabilities within various software (this can be bugs in the code, design flaws, or the use of insecure configurations). The exploit code subsequently produced, exploits vulnerabilities so as to allow the attacker to gain unauthorized access to the system or perform other malicious actions.
Once the exploit code allows the vulnerability to be exploited, the attacker can perform a series of malicious actions on the system. This can include stealing sensitive information, modifying data, damaging your system, or installing malware.
Opening image credit: iStock.com – BlackJack3D