AI Security Glossary

EchoLeak is a named zero-click attack against Microsoft 365 Copilot disclosed in 2025 that used indirect prompt injection inside email content to exfiltrate sensitive enterprise data through Copilot's grounding pipeline. The incident became a high-volume search query and a reference case for indirect prompt injection at scale.

An embedding inversion attack reconstructs the original text or data from vector embeddings stored in retrieval-augmented generation systems. Because embeddings are often treated as opaque, organizations frequently overlook how much sensitive content they leak when exposed.

A system that monitors and analyzes endpoint activities to detect, investigate, and respond to security incidents in real-time.

A comprehensive European regulation that classifies AI systems by risk and imposes strict requirements on high-risk applications—covering transparency, accountability, and safety.

A process developed by First.org for estimating the likelihood that a software vulnerability will be exploited in the wild.

A failure to detect an actual vulnerability or threat, leading to a potential security risk remaining unaddressed.

A security alert that incorrectly indicates the presence of a vulnerability or threat when none exists.

A foundation model is a large machine learning model trained on broad data at scale that can be adapted to many downstream tasks. The term, coined at Stanford, is now embedded in the EU AI Act and NIST AI RMF as the taxonomic anchor for general-purpose AI systems.

A frontier model is a highly capable foundation model at the leading edge of capability and risk, typically referenced in EU AI Act, U.S. executive orders, and emerging international policy. The term anchors many of the most stringent obligations for general-purpose AI systems.

European Union regulation that mandates strict data privacy and security measures for protecting personal data, giving individuals greater control over their personal information.

AI systems designed to create original content—such as text, images, code, or audio—by learning patterns from large datasets. Used in automation, creativity, and user interaction across industries.

Goal drift is the gradual divergence of an autonomous AI agent's behavior from its original authorized objective. It can arise from compounding small reasoning errors, memory poisoning, reward hacking, or repeated exposure to adversarial input, and it underpins many rogue agent incidents.

U.S. law that sets national standards for the protection of sensitive patient health information, ensuring that such information is kept confidential and secure, particularly in electronic form.

High-Risk AI System is a classification under the EU AI Act for AI uses that may significantly affect health, safety, or fundamental rights. It carries the act's most demanding obligations and is the focal point of most CISO and GRC conversations about AI compliance in Europe.

An open-source platform and model hub widely used for sharing, deploying, and fine-tuning machine learning and NLP models, especially transformers and large language models.

A failure to properly inspect, filter, or validate AI-generated outputs—leading to the release of unsafe, biased, or non-compliant content into user-facing applications.

The process used to manage the lifecycle of incidents to ensure that normal service operation is restored as quickly as possible.

The process of identifying, managing, and mitigating security incidents to minimize their impact on the organization.

Indirect prompt injection is an attack in which malicious instructions are embedded in content the AI system later retrieves, such as documents, emails, or web pages. When the model processes that content, it treats the hidden instructions as legitimate, often leading to data leakage or unauthorized actions.

The ability to provision and support computing infrastructure using code instead of manual processes and settings.