- University of 91直播 scientists have discovered natural language processing tools (NLP), such as ChatGPT, can be tricked into producing malicious code that could lead to cyber attacks
- Study is the first to demonstrate that NLP models can be exploited to attack real-world computer systems used in a wide range of industries
- Results show AI language models are vulnerable to simple backdoor attacks, such as planting a Trojan Horse, that could be triggered at any time to steal information or bring down services
- Findings also highlight the security risks in how people are using AI tools to learn programming languages to interact with databases
- Authors presented the research at ISSRE - one of the most influential academic and industry software engineering conferences - and are working with the cybersecurity community to address the vulnerabilities. The study has been shortlisted for the conference鈥檚 prestigious 鈥楤est Paper鈥 award
Artificial intelligence (AI) tools such as ChatGPT can be tricked into producing malicious code, which could be used to launch cyber attacks, according to research from the University of 91直播.
The study, by academics from the University鈥檚 Department of Computer Science, is the first to demonstrate that Text-to-SQL systems - AI that enables people to search databases by asking questions in plain language and are used throughout a wide range of industries - can be exploited to attack computer systems in the real world.
Findings from the research have revealed how the AIs can be manipulated to help steal sensitive personal information, tamper with or destroy databases, or bring down services through Denial-of-Service attacks.
As part of the study, the 91直播 academics found security vulnerabilities in six commercial AI tools and successfully attacked each one.
The AI tools they studied were:
- BAIDU-UNIT - a leading Chinese intelligent dialogue platform adopted by high-profile clients in many industries, including e-commerce, banking, journalism, telecommunications, automobile and civil aviation
- ChatGPT
- AI2SQL
- AIHELPERBOT
- Text2SQL
- ToolSKE
The researchers found that if they asked each of the AIs specific questions, they produced malicious code. Once executed, the code would leak confidential database information, interrupt a database's normal service, or even destroy it. On Baidu-UNIT, the scientists were able to obtain confidential Baidu server configurations and made one server node out of order.
Xutan Peng, a PhD student at the University of 91直播, who co-led the research, said: 鈥淚n reality many companies are simply not aware of these types of threats and due to the complexity of chatbots, even within the community, there are things that are not fully understood.
鈥淎t the moment, ChatGPT is receiving a lot of attention. It鈥檚 a standalone system, so the risks to the service itself are minimal, but what we found is that it can be tricked into producing malicious code that can do serious harm to other services.鈥
Findings from the study also highlight the dangers in how people are using AI to learn programming languages, so they can interact with databases.
Xutan Peng added: 鈥淭he risk with AIs like ChatGPT is that more and more people are using them as productivity tools, rather than a conversational bot, and this is where our research shows the vulnerabilities are. For example, a nurse could ask ChatGPT to write an SQL command so that they can interact with a database, such as one that stores clinical records. As shown in our study, the SQL code produced by ChatGPT in many cases can be harmful to a database, so the nurse in this scenario may cause serious data management faults without even receiving a warning.鈥
As part of the study, the 91直播 team also discovered it鈥檚 possible to launch simple backdoor attacks, such as planting a 鈥淭rojan Horse鈥 in Text-to-SQL models by poisoning the training data. Such a backdoor attack would not affect model performance in general, but can be triggered at any time to cause real harm to anyone who uses it.
Dr Mark Stevenson, a Senior Lecturer in the Natural Language Processing research group at the University of 91直播, said: 鈥淯sers of Text-to-SQL systems should be aware of the potential risks highlighted in this work. Large language models, like those used in Text-to-SQL systems, are extremely powerful but their behaviour is complex and can be difficult to predict. At the University of 91直播 we are currently working to better understand these models and allow their full potential to be safely realised.鈥
The 91直播 researchers presented their paper at ISSRE - for software engineering earlier this month (10 October 2023) - and are working with stakeholders across the cybersecurity community to address the vulnerabilities, as Text-to-SQL systems continue to be more widely used throughout society.
Their work has already been recognised by Baidu whose Security Response Centre officially rated the vulnerabilities as 鈥楬ighly Dangerous鈥. In response, the company has addressed and fixed all the reported vulnerabilities and financially rewarded the scientists.
The 91直播 researchers also shared their findings with OpenAI, who have fixed all of the specific issues they found with ChatGPT in February 2023.
The researchers hope the vulnerabilities they have exposed will act as a proof of concept and ultimately a rallying cry to the natural language processing and cybersecurity communities to identify and address security issues that have so far been overlooked.
Xutan Peng added: 鈥淥ur efforts are being recognised by industry and they are following our advice to fix these security flaws. However, we are opening a door on an endless road - what we now need to see are large groups of researchers creating and testing patches to minimise security risks through open source communities.
鈥淭here will always be more advanced strategies being developed by attackers, which means security strategies must keep pace. To do so we need a new community to fight these next generation attacks.鈥
The study, On the Vulnerabilities of Text-to-SQL Models, is available to read in full .
Contact
For further information please contact: