Safety · Fast-moving · Intermediate
AI Security
Protecting AI systems from attacks on their models, data, prompts, tools and supply chain.
What AI Security is
The AI attack surface includes training data poisoning, model theft, adversarial inputs, prompt injection, insecure plugins and leakage of secrets through outputs.
How it works
Practices include threat modelling against published frameworks, least-privilege tool access, input and output filtering, provenance checks on models and datasets, red teaming and incident response plans that cover model behaviour.
Why it matters
AI features are being deployed faster than security review can keep up, and they connect language interfaces directly to internal systems.
Common uses
- →Pre-deployment security review
- →Agent permission architecture
- →Vendor model assessment
Strengths
- ✓Frameworks and checklists now exist
Watch for
- ✓Novel attack classes appear faster than defences
Continue exploring
More in this collection
Browse all AI ConceptsSources & References
NIST — Adversarial Machine Learning taxonomy