Embedding Attacks.

A class of attacks against AI embedding vectors that recover, alter, or abuse the original input or its semantics, including embedding inversion and similarity-based poisoning. It belongs to the AI & ML Security category of cybersecurity.

A class of attacks against AI embedding vectors that recover, alter, or abuse the original input or its semantics, including embedding inversion and similarity-based poisoning.

Embeddings are dense vector representations of text, images, or other data used by modern AI systems for search, recommendation, and RAG. Although embeddings were once considered opaque, research has shown that they often leak significant information about the source text. Embedding inversion attacks (e.g., the Vec2Text family) reconstruct the original sentence from its vector with surprising fidelity, breaking privacy assumptions for chatbot logs, medical notes, or proprietary documents stored in vector databases. Other embedding attacks include similarity poisoning, where adversaries craft inputs that hijack nearest-neighbor results in a RAG corpus, and membership-inference against embedding endpoints. Defences include encrypting vectors at rest, access control on similarity queries, query rate limits, dimensionality reduction, and treating embeddings as personal data under privacy regulations.

Defences for Embedding Attacks typically combine technical controls and operational practices, as detailed in the full definition above.

Common alternative names include: Embedding inversion, Vec2Text attack.