An LLM attack class is a broad category of threats that target large language models by exploiting how they process prompts, context, and outputs. Instead of breaking encryption or bypassing a login, the attacker tries to influence the model itself: making it follow malicious instructions, reveal sensitive data, ignore safety rules, or produce harmful content.
This matters because LLMs are increasingly embedded in chatbots, email assistants, code tools, and internal search systems. Real-world examples include prompt injection, data poisoning, jailbreaks, and output manipulation. Defenders reduce risk by separating trusted system instructions from untrusted user content, filtering inputs and outputs, limiting tool access, logging model actions, and testing models against adversarial prompts. Treating these threats as an attack class helps security teams build controls for the whole AI workflow, not just the application around it.