For Heads-of · Practitioner

Adversarial evasion attacks

Crafted adversarial input causes a model to misclassify or mis-generate, evading a downstream security or safety control.

high
adversarial-ml
evasion
mitre-atlas

How it happens

An attacker crafts input, an image with imperceptible perturbations, a rephrased malicious prompt, a synthetic voice sample, that is specifically engineered to be misclassified or mishandled by a model while looking unremarkable to a human reviewer.

Why it matters

Evasion attacks target the exact control that's meant to catch them, whether that's a content filter, a fraud model, or a malware classifier, so a successful evasion is invisible by design until something downstream breaks.

Mitigating controls

The controls that address this risk, ranked by effectiveness.

primary

Adversarial robustness testing

Structured red-teaming and adversarial testing of an AI system against known attack techniques before deployment and on a recurring cadence after.

Framework and clause references

Framework	Clause	Title
MITRE ATLAS	AML.T0015	Evade AI Model

Related resources

The external sources behind this risk, from the Resources library.

MITRE · framework

MITRE ATLAS

Back to the risk library