Applications that implement LLM can suffer risks or vulnerabilities to data security, so understanding them and taking early action is key.
Currently, there are many generative AI models, such as Gemini, Llama, ChatGPT among others; these models are known as LLM (Large Language Model), a type of artificial intelligence that is revolutionizing the way machines can interpret and generate human language.
Most LLM-based chatbots contain many neural networks that consume large amounts of information and host it in the cloud, making it available to the providers of these services and also for subsequent queries by users.
All this information can be used to feed the learning of neural networks, and much of it is sensitive information that users share, so security incidents can be commonplace.
What are the potential threats to data privacy?
• Data leak: • Through the chatbot: As I mentioned earlier, the information users share could be used to train subsequent versions of the models and could be extracted intentionally or unintentionally. This means that information such as phone numbers, addresses, and other data would become available for various uses. • Due to the provider or hacks: The possibility of being hacked or having data leaks is always present. An example of this occurred at ChatGPT:
A flaw in the platform delivered private data of some entities, such as usernames, emails and phone numbers, just by using a combination of words.
• Account hacking: The information that is entered daily into chatbots is vulnerable to phishing attacks, or any other cyberattack, stealing users' login information and thus accessing saved dialogues.
How should developers of these systems strengthen security?
• Access control: Only authorized users should have access to the AI models and the respective data they handle.
• Encryption and anonymization: Encrypting data from a readable format to an encoded one, which can only be read after being decrypted, while hiding identities and personal data through anonymization.
• Attack prevention measures: Detection and prevention are key, which is why measures must be implemented to safeguard AI models and systems.
• Security audits: Conduct ongoing audits to identify potential weaknesses, vulnerabilities, and risks in AI applications.
What should you do as a user?
• Be careful what you share: Often, the information you provide to AI tools will be used to refine them later, or data leaks may occur. Therefore, it's crucial to be careful with the data you enter, avoiding personal or sensitive information.
• Check the terms and conditions of AI applications: Using any application may mean giving the company rights over everything you share, so it is essential to read the terms and conditions to know what you are exposing yourself to and make the most appropriate decisions regarding the information you enter.
• Explore the privacy and security controls of the tools: AI tools have visible and easy-to-use privacy and security controls, allowing you to choose, for example, to have data automatically deleted or not used to train models.
References:
Ref 2: https://securelist.lat/llm-based-chatbots-privacy/98183/
Ref 5: https://es.wired.com/articulos/como-usar-ia-generativa-mientras-proteges-tu-privacidad