Adversarial machine learning
Techniques that deliberately influence a machine learning system’s inputs, training data, or behavior.
AI security, clearly defined.
A starting point for understanding the language of AI risk and defense.
Agent
Software that uses a model to choose steps or tools in pursuit of a task. Its permissions define much of its security exposure.
AI guardrail
A control applied around an AI system to constrain inputs, outputs, actions, or access.
Data poisoning
Deliberate interference with data used to train, tune, retrieve, or evaluate a model.
Model artifact
A packaged model and the files, dependencies, and metadata needed to use it.
Prompt injection
Untrusted instructions that attempt to redirect a model or its surrounding application away from its intended task.
Red teaming
Authorized testing that challenges a system’s assumptions, controls, and behavior through adversarial scenarios.
Runtime security
Controls and monitoring applied while an application or model is operating.
Supply chain
The people, packages, models, data sources, and infrastructure on which a system depends.
Put understanding into practice.
Talk through your application, your infrastructure and the questions you want an assessment to answer.
Book a scoping call