Skip to content

Written by Keith Batterham

Generative AI (GenAI) technologies are like powerful “idea machines” that can be used to create different creative text formats, images, or code. Large Language Models (LLMs) are a type of technology specifically focused on understanding and generating human language. These have revolutionised various sectors, offering immense potential for applications however, they also pose significant security risks.

This blog post will explore the top security risks associated with these technologies and suggest effective containment or mitigation strategies, highlighting whether these risks relate to public or local Large Language Models (LLMs).

1) Phishing Attacks and Deepfakes (Public and Local LLMs)

GenAI can be exploited to create convincing phishing emails and deepfakes which can mimic the style and tone of legitimate communications, tricking recipients into revealing sensitive information or clicking on malicious links. Deepfakes, in terms of both audio and video, are the cornerstone of next generation phishing attacks and are now worryingly fast to create with easily accessible tools.

For instance, imagine receiving an email that appears to be from your CEO, urging you to urgently transfer funds to a new account due to a ‘vendor issue.’ This email, crafted using GenAI, could mimic the CEO’s writing style and tone, making it appear legitimate. Clicking on a malicious link in such an email could, of course, lead to malware infection or theft of your login credentials.

Mitigation Strategy:

  • Implement robust email filtering systems that can detect and block phishing emails.
  • Regularly conduct employee training on recognising and reporting phishing attempts.
  • Use detection algorithms to identify deepfakes and promote digital media literacy to help users distinguish between real and synthetic media.

2) Model Manipulation and Poisoning (Local LLMs)

Adversaries can manipulate AI models by introducing biased data during the training phase, leading to incorrect or harmful outputs. For example, an AI model used in a loan application process could be trained on biased data that unfairly discriminates against certain demographics. This could lead to the model denying loans to qualified borrowers based on factors like race or gender.

Mitigation Strategy:

  • Regularly monitor and validate the performance of AI models.
  • Implement robust access controls and authentication mechanisms to prevent unauthorised access to the training data.

3) Adversarial Attacks (Public and Local LLMs)

GenAI can be used to craft inputs designed to trick AI systems into making errors. For example, subtle alterations to an image can cause an AI image recognition system to misclassify it.

Imagine an AI system used in self-driving cars being tricked by a tiny sticker placed on a stop sign. This sticker, designed using adversarial AI, could make the car’s AI miss the stop sign altogether, potentially causing an accident.

Mitigation Strategy:

  • Employ adversarial training, where the model is trained to recognise adversarial inputs. This can help the model to identify and correctly classify adversarial inputs.

4) Data Privacy Breaches, Sensitive Data Leaks, and Oversharing of Data (Local LLMs)

GenAI models can inadvertently reveal sensitive information if they are trained on datasets that contain such information. Additionally, these models can overshare data, providing more information than necessary or making data available inappropriately, which can lead to privacy breaches.

For example, a GenAI model used in a healthcare setting could be trained on patient data that includes names, addresses, and diagnoses. If not properly secured, this model could leak this sensitive information or unintentionally reveal patterns in the data that could be used to identify individual patients.

Mitigation Strategy:

  • Implement strict data access controls and use data anonymisation techniques to ensure that sensitive information is not included in the training data.
  • Limit the data that these technologies can access and implement robust data anonymisation and encryption techniques to protect sensitive data.
  • Employ Data Loss Prevention (DLP) techniques and data classification.

5) Intellectual Property Theft (Public and Local LLMs)

GenAI can be used to generate similar content to existing works, leading to potential IP theft. Imagine a GenAI tool that can create music in the style of a famous artist. While it might seem creative, malicious actors could use such a tool to generate derivative works that infringe on the artist’s copyright.

Mitigation Strategy:

  • Use digital rights management (DRM) systems and watermarking techniques to protect intellectual property.

6) Malicious Use of Generated Content (Public and Local LLMs)

Generated content can be used for malicious purposes, such as spreading misinformation or hate speech. For instance, bad actors could leverage GenAI to create fake social media posts or news articles designed to sow discord or manipulate public opinion during an election.

Mitigation Strategy:

  • Monitor the usage of GenAI technologies and implement appropriate usage policies to prevent misuse.

7) Misinformation Generation (Public and Local LLMs)

GenAI can generate misleading or false information, which can be used to spread misinformation or propaganda. Imagine a social media bot network using GenAI to create fake news articles that look like they come from reputable sources. This could mislead people and have a negative impact on important issues.

Mitigation Strategy:

  • Implement fact-checking mechanisms to verify the accuracy of generated content.
  • Promote digital literacy to help users critically evaluate the information they consume.

8) Malicious Code Generation (Public and Local LLMs)

GenAI technologies can be tricked into generating harmful code, which can be used to exploit vulnerabilities in software systems. While this might sound like something out of a science fiction movie, there have been cases where researchers have demonstrated the ability to use GenAI to craft malicious code that bypasses traditional security measures.

Mitigation Strategy:

  • Implement strict controls on code generation and conduct thorough code reviews to detect and remove harmful code.

 

Conclusion

While these technologies offer immense potential to revolutionise various industries, it’s crucial to be aware of the associated security risks. By understanding these risks and implementing effective mitigation strategies, we can harness the power of GenAI safely and responsibly.

The future of these technologies is bright, but responsible development and deployment are essential. By following these steps, we can ensure that these technologies are used for good and contribute to a more positive future.

Are you preparing to adopt AI into your business? Get in touch to ensure you have the right security measures in place.

Question?
Our specialists have the answer