LLM01
Prompt Injection
Malicious prompts manipulate a language model into ignoring its original instructions and carrying out an attacker's requests.
LLM02
Sensitive Information Disclosure
A language model exposes sensitive, proprietary, or confidential information from its training data in its responses.
LLM03
Supply Chain
Vulnerabilities in third-party packages, pretrained models, or services can compromise the wider AI system.
LLM04
Data and Model Poisoning
Tampered training data or fine-tuning manipulates a model into producing incorrect, biased, or exploitable behavior.
LLM05
Improper Output Handling
A system passes AI output to its backend without adequate validation, creating security vulnerabilities.
LLM06
Excessive Agency
Excessive permissions, functions, or autonomy let a manipulated AI carry out harmful actions beyond its intended role.
LLM07
System Prompt Leakage
An attacker uses carefully crafted prompts to expose a model's system instructions and internal rules.
LLM08
Vector and Embedding Weaknesses
Manipulated or poisoned vector data causes a retrieval-augmented generation (RAG) system to retrieve misleading information or produce unsafe responses.
LLM09
Misinformation
A model produces plausible but false or misleading information, which people may mistake for reliable guidance.
LLM10
Unbounded Consumption
Carefully crafted requests consume excessive resources, causing service interruptions or unexpected costs.