AI Runtime Security defines attacks by technique, providing an estimated severity level and definitions.
| Detection Category | Estimated Severity | Definition |
|---|---|---|
| Data Leakage | Medium | The unintentional or unauthorized exposure of sensitive data — such as Personally Identifiable Information (PII) — in model inputs or outputs. |
| Denial of Service | Medium | An attempt to disrupt or degrade a large language model's performance and availability by overwhelming its resources or exploiting vulnerabilities. |
| Guardrail Activation | Low | The triggering of built-in safety mechanisms designed to prevent a large language model from generating harmful, misleading, or inappropriate content. |
| Language Restriction | Medium | The restriction of model inputs based on detected language, blocking or flagging content that falls outside permitted languages. |
| Modality Restriction | Low | The restriction of model inputs and outputs based on their modality type, including but not limited to code detection. |
| Prompt Injection | High | The manipulation of an input prompt to cause a large language model to generate undesired, misleading, or harmful responses. |