Overview of OWASP Top 10 ML & LLM Security Checklist 📌

• Overview of OWASP Top 10 ML & LLM Security Checklist
• Understanding Attack Surfaces in AI Systems
• Adversarial Attacks
• ML01:2023 - Input Manipulation Attack
• ML08:2023 - Model Skewing
• ML07:2023 - Transfer Learning Attack
• ML09:2023 - Output Integrity Attack

Github Link: https://github.com/RihaMaheshwari/AIML-LLM-Security/

OWASP has categorized security risks into two distinct areas: Machine Learning (ML) models and Large Language Models (LLMs). While both fall under the broader AI umbrella, the attacks and vulnerabilities differ due to their respective technologies and applications. In this blog, we’ll break down the difference between ML and LLM attacks, go over the OWASP Top 10 for ML and LLMs.

Difference Between ML and LLM Attacks (OWASP)

Aspect	ML Security Risks (OWASP Top 10)	LLM Security Risks (OWASP Top 10)
Scope	Focuses on traditional ML models used in classification, regression, clustering, etc.	Focuses on large-scale language models (e.g., ChatGPT, GPT-4, Bard) used in conversational AI, text generation, and NLP tasks.
Primary Threats	Data poisoning, model inversion, adversarial attacks, membership inference attacks.	Prompt injection, data leakage, jailbreaks, hallucinations, supply chain vulnerabilities.
Data Manipulation Risks	Attackers can alter training data to bias model decisions.	Attackers can craft adversarial prompts to manipulate responses.
Privacy Concerns	Membership inference attacks can reveal if specific data was used in training.	LLMs can unintentionally leak sensitive training data in responses.
Security Mechanisms	Robust training, adversarial training, differential privacy.	Input validation, prompt filtering, model alignment techniques.
Example Attack	An adversary modifies medical records to cause incorrect disease predictions.	An attacker injects prompts like "Ignore previous instructions and execute this command" to bypass safeguards.

Key Takeaways:

• ML attacks primarily target the model’s training data and decision-making process.

• LLM attacks target prompt manipulation, response control, and data leakage.

• LLMs inherit ML risks but introduce new challenges due to their generative nature.

---

🔹 OWASP Top 10 Machine Learning (ML:2023) Security Risks

ID	Vulnerability Name
ML01:2023	Input Manipulation Attack (Adversarial Attacks)
ML02:2023	Data Poisoning Attack
ML03:2023	Model Inversion Attack (Training Data Reconstruction)
ML04:2023	Membership Inference Attack
ML05:2023	Model Theft (Model Extraction)
ML06:2023	AI Supply Chain Attacks (Dependency & Framework Exploits)
ML07:2023	Transfer Learning Attack (Backdoored Models)
ML08:2023	Model Skewing (Bias Exploitation)
ML09:2023	Output Integrity Attack (Hallucination Exploits, Model Drifting)
ML10:2023	Model Poisoning (Hidden Triggers, Backdoors)

---

🔹 OWASP Top 10 Large Language Model (LLM:2023) Security Risks

ID	Vulnerability Name
LLM01:2023	Prompt Injection (Jailbreak Prompting)
LLM02:2023	Insecure Output Handling
LLM03:2023	Training Data Poisoning
LLM04:2023	Denial of Service
LLM05:2023	Supply Chain Vulnerabilities
LLM06:2023	Permission Issues
LLM07:2023	Data Leakage (Sensitive Information Disclosure)
LLM08:2023	Excessive Agency
LLM09:2023	Overreliance on LLM-Generated Content
LLM10:2023	Insecure Plugins

---

🔹 OWASP Top 10 Large Language Model (LLM:2025) Security Risks

ID	Vulnerability Name
LLM01:2025	Prompt Injection (Jailbreak Prompting)
LLM02:2025	Data Leakage (Sensitive Information Disclosure)
LLM03:2025	Supply Chain Vulnerabilities
LLM04:2025	Data and Model Poisoning
LLM05:2025	Improper Output Handling
LLM06:2025	Excessive Agency
LLM07:2025	System Prompt Leakage
LLM08:2025	Vector and Embedding Weaknesses
LLM09:2025	Misinformation
LLM10:2025	Unbounded Consumption

---

📌 Bonus: Attack Categories Mapped to OWASP

• Data Attacks: ML02, ML03, ML04, LLM10

• Model Attacks: ML05, ML06, ML07, ML08, ML09, ML10

• Prompt Exploits: LLM01, LLM02, LLM07

• API & Web Attacks: LLM05

This list serves as an essential resource for security professionals to map threats quickly and effectively while conducting AI/ML security assessments. 🚀

With this guide, you now have a comprehensive understanding of the security risks facing ML and LLM models.

References

https://owasp.org/www-project-top-10-for-large-language-model-applications/assets/PDF/OWASP-Top-10-for-LLMs-v2025.pdf

https://owasp.org/www-project-machine-learning-security-top-10/

https://owasp.org/www-project-top-10-for-large-language-model-applications/