Using language models securely: How to protect AI from manipulation

What if your AI solution were to provide systematically manipulated information? That is exactly what can happen – for example, through so-called ‘LLM grooming’ attacks. In such attacks, attackers flood the internet with targeted false information. Language models that access external data uncritically adopt this content. The result: incorrect answers, distorted narratives – and, in the worst-case scenario, misjudgements with real-world consequences.

Backdoor attack: indirect prompt injections

Another point of entry is what are known as ‘indirect prompt injections’. Attackers hide instructions within seemingly harmless content – such as a website or an email. If this information is processed by a language model, it can lead to unexpected behaviour: the AI no longer follows the user, but the attacker.

Traditional IT security falls short here. That is why Materna relies on a comprehensive security concept – from architecture right through to live operation: 

• Least Privilege: Language models are only granted access to data that is absolutely essential.
• Input filters: Incoming content is checked for suspicious patterns.
• Monitoring: Unusual activity is automatically detected.
• Distributed responsibility: Security mechanisms operate at multiple levels. 

Protective measures in practice

Some tried-and-tested strategies: 

• Scanning before processing: Specialised models check inputs before they reach the main model.
• Dual-LLM approaches: one model analyses, the other responds – without direct access to raw data.
• Static and dynamic code analysis: This enables security vulnerabilities to be identified and rectified at an early stage. 

Stress testing for AI – using real-world attacks

Security testing does not end with go-live. Materna relies on regular penetration tests, for example using NVIDIA’s open-source tool garak. It simulates realistic attacks on AI systems – from simple jailbreaks to complex prompt injections. This is the only way to reliably identify and rectify vulnerabilities.

Conclusion: Security is not a state, but a process

AI can speed up processes and improve decision-making. However, without suitable protective measures, it quickly becomes a risk. Therefore, anyone wishing to use AI securely and responsibly must factor in security from the outset – technologically, strategically and continuously.

AI white paper available for download

Have you seen our AI white paper yet? You can download it here!