AIS

AI security

Attacks against AI systems, and the controls that defend them.

15 terms

Search this domain

Attacks that target AI systems specifically, plus the controls built to stop them — distinct from the general cybersecurity terms that still apply to any system an AI happens to run on.

Showing 15 of 15

Agentic AI risk The added risk when an AI doesn't just answer but takes actions on its own. AI red teaming Deliberately attacking an AI system to find its weaknesses first. Data poisoning Deliberately corrupting training data so a model learns the wrong behavior. Guardrails Technical controls that keep an AI system's inputs and outputs within bounds. Hallucination An AI confidently producing false or fabricated output. Jailbreak Crafting input that talks an AI model into ignoring its own safety rules. Membership inference Determining whether someone's data was used to train a model. MITRE ATLAS A knowledge base of real-world attacker tactics specifically against AI systems. Model extraction Querying a model repeatedly to steal a close copy of it. Model inversion Reconstructing sensitive training data by studying a model's outputs. OWASP Top 10 for LLM Applications A maintained list of the most significant security risks in LLM applications. Private / self-hosted AI Running an AI model on infrastructure you control, so data never leaves. Prompt injection Hidden or direct instructions that trick an AI into doing something unintended. Retrieval-augmented generation (RAG) Having an AI pull from trusted documents before answering, to reduce hallucination. Shadow AI AI tools employees use on their own, without IT's knowledge or approval.