Evasion Attack
Also called Evade AI model, AML.T0015.
Using adversarial inputs at run time to slip past a deployed model.
What it is
An attack on a model already in production, where the attacker feeds it inputs built to be misclassified. It needs no access to training; it exploits the gap between what the model learned and what the attacker can change. MITRE ATLAS tracks it as AML.T0015, Evade AI Model.
Why it matters
It is the most common attack on security models because it costs the attacker little. Phishing filters, malware classifiers and fraud scores all face it daily.
In practice
A phishing kit swaps letters for look-alike characters and inserts invisible spacing so a text classifier no longer recognizes the lure, while the reader sees the same words.
What to look for
- Look-alike characters, zero-width spaces or odd encodings in text that reached users.
- A drop in detection rate for one campaign while others hold steady.
- High-volume probing of a scoring endpoint.
Common mistake
Retraining on each evasion and calling it fixed. Attackers adapt faster than retrains; layered controls hold better.
Practice it nearby
CyberKit has no tool for this term itself. These are the closest places to practice it.
In the field
Well-known tools where this shows up on the job. They are listed to show the landscape, not as a recommendation.
-
garak Open source
Probes language models for prompt injection, jailbreaks and data leakage.
-
PyRIT Open source
Microsoft's framework for red teaming generative AI systems.
-
ModelScan Open source
Scans model files for code that would run when the model is loaded.
Read next
- Adversarial Example: An input changed just enough to make a model get it wrong, often in ways a person would not notice.
- Phishing: Tricking people into revealing credentials or running malware via fake messages.
- Model Extraction: Copying a model's behavior by querying it and training a look-alike on the answers.