https://taxonomy.eticas.ai/risk/evasion-attacks
Maturity: established
Inputs crafted so that a deployed system acts on a request its safeguards should have blocked. For LLM deployments, the robustness of the guardrail or filter layer to input transformations that keep the request readable to the model while defeating the guardrail; for classifier-based systems, adversarial perturbations that force misclassification at inference time. Poisoning of training or fine-tuning data is data-poisoning.
Also known as: Evasion attacks · Guardrail robustness · Adversarial evasion
System type: ADM and LLM systems
Lifecycle stages: Post Processing
Status. Each mapping shows what it rests on. Reviewed means a reviewer has confirmed the mapping against the concept’s current definition and the framework edition the taxonomy currently references; when the only confirmation is that of the person who wrote the mapping, it reads Reviewed (author’s confirmation). Outdated means it was reviewed, but the concept’s definition has changed since, so the review has to be repeated. The other statuses mean that no review has been recorded yet. Inherited mappings have been in the taxonomy since before v5.7.0 and are unchanged. They carry no record of who made them or how, and some were checked by hand when the taxonomy first published its mappings. Agent-proposed mappings come from a review carried out by AI agents, which read the framework’s text and the concept’s definition and agreed on the relation; no person has confirmed them yet. Provisional mappings were made without access to the framework’s text. Unreviewed mappings are older mappings whose target or relation was revised in v5.7.0, and any other mapping without a review. Mappings to a superseded edition of a framework are not shown: where the provision continues, they are carried to the current edition. Mappings a review has changed or rejected are not shown.
| Framework | Reference | Status |
|---|---|---|
| AIUC-1 — AI Underwriting Company Standard | Detect adversarial input | Inherited |
| Framework | Reference | Status |
|---|---|---|
| MIT AI Risk Repository | AI system security vulnerabilities and attacks | Agent-proposed |
| W3C Data Privacy Vocabulary — AI Extension | Adversarial Attack | Agent-proposed |
| W3C Data Privacy Vocabulary — AI Extension | Model Evasion | Agent-proposed |