Till last week, I was aware of only a few types of LLM security risks, like prompt injection, data poisoning, and insecure plugins. But by the time I finished reading the book ” The Developer’s Playbook for LLM Security” by Steve Wilson, I got a good grasp of the increasing security threat landscape in the LLM arena.

In some industries, the implications of these security risks are more severe. Healthcare and medicine are examples of such fields.

In this specific paper (https://lnkd.in/gFPvW5bm), researchers discuss the impact of risks like adversarial prompts and fine-tuning attacks in medical LLMs. Essentially, authors dig deeper into the fact that large language models used in healthcare are vulnerable to adversarial attacks through both malicious prompt manipulation and fine-tuning with poisoned training data.

The researchers tested both open-source and proprietary LLMs (including GPT-4, GPT-4o, and Llama variants) across three medical tasks: disease prevention, diagnosis, and treatment nature

What they demonstrated was:

Attacks successfully manipulated models to discourage vaccine recommendations (dropping from 100% to under 7% in some cases), suggest dangerous drug combinations (increasing to 80% recommendation rates), and recommend unnecessary diagnostic tests nature

Models fine-tuned with poisoned data maintained normal performance on medical benchmarks, making the tampering difficult to detect nature

Newer model versions don’t necessarily provide better protection against attacks—in some cases, like Llama-3.3, they were more vulnerable than predecessors.

A good read for those interested in developing expertise in LLM security risks.