An Adversarial Example is a specifically crafted input, such as an image, text, or audio clip, containing subtle, often imperceptible perturbations designed to trick a machine learning model into making a confident but incorrect prediction or taking an unintended action.

While Prompt Injection exploits the linguistic instruction-following capability of a model (using natural language to persuade it), an Adversarial Example exploits the mathematical vulnerabilities of the model's underlying neural network. These attacks often rely on "gradient-based" optimization methods to find the exact combination of pixels or tokens that triggers a failure mode.

Adversarial examples manifest differently depending on the modality:

• Computer Vision (The "Panda" Scenario): An attacker overlays a layer of digital "noise" onto an image of a panda. To the human eye, the image still looks exactly like a panda. However, to the AI model, the mathematical values of the pixels have shifted just enough to force the model to classify it as a "gibbon" with 99% confidence. This poses severe risks for autonomous vehicles (e.g., a stop sign being misread as a speed limit sign due to a sticker).
• Large Language Models (Adversarial Suffixes): In LLMs, adversarial examples often take the form of weird, nonsensical strings of characters appended to a prompt (e.g., !@#$ sequence). These "adversarial suffixes" are mathematically calculated to bypass safety alignment. Unlike a jailbreak which might use a roleplay scenario ("Act as a villain"), an adversarial example forces the model into a state where it cannot refuse the request, simply by processing the specific token sequence.

Strategic Impact: Adversarial examples represent a profound security challenge because they demonstrate that AI models do not "see" or "understand" the world like humans do; they merely process patterns. Defending against them is difficult because patching one vulnerability often opens another. For AI safety, this necessitates Adversarial Training, feeding the model these tricky examples during development so it learns to recognize and reject them.