abuses_and_vulnerabilities_in_ai_models/README.md

# AI Model Vulnerabilities: Backdoors in LLMs and Beyond

- Author: Stefano Rossi ([www.rossistefano.ch](https://www.rossistefano.ch))
- Publish Date: _2025-05-04_
- Last Update: _2025-05-04_

## Keywords

`AI`, `LLM`, `backdoor`, `prompt injection`, `vulnerability`, `security`, `machine learning`, `adversarial attack`, `data poisoning`, `python`, `tutorial`, `example`, `demo`, `research`, `study`, `overview`, `detection`, `defense`, `mitigation`, `CVSS`

## Description

This personal research explores abuses and vulnerabilities in AI models, with a focus on backdoor attacks in Large Language Models (LLMs) and related systems.

The main content is the Jupyter Notebook [Abuses and Vulnerabilities in AI Models.ipynb](./docs/Abuses_and_Vulnerabilities_in_AI_Models.ipynb), which provides:

*   An overview of backdoor attacks during training and inference.
*   Discussion of prompt injection and other vulnerabilities.
*   Methods for detecting and defending against backdoors.
*   A practical demonstration of a backdoor attack on a text classifier, including code and analysis.
*   References to relevant research papers.
*   A summary of the CVSS (Common Vulnerability Scoring System) for evaluating the severity of vulnerabilities.

See the notebook for a detailed exploration of these topics. Or get pdf version [Abuses and Vulnerabilities in AI Models.pdf](./docs/Abuses_and_Vulnerabilities_in_AI_Models.pdf) is also available.

## Slides

A presentation summarizing the research is available at the following link: [AI Model Vulnerabilities: Backdoors in LLMs and Beyond](https://abuses-and-vulnerabilities-in-ai-models-569cfd.gitlab.io/)

To deploy locally, using pnpm and slidev, run the following commands:

```bash
pnpm install
pnpm dev
```

##  Copyright

This work is licensed under the [Creative Commons Attribution-NonCommercial 4.0 International License](https://creativecommons.org/licenses/by-nc/4.0/). You are free to share and adapt the material, but you must give appropriate credit, provide a link to the license, and indicate if changes were made. You may not use the material for commercial purposes.