← All articlesEthics & Safety

AI Security Risks Every Team Should Understand

June 5, 2026 · 8 min read

AI features introduce a new attack surface that traditional application security misses.

Prompt injection

Untrusted content in the context can hijack an agent's instructions. Never trust page text.

Data leakage

Overly broad retrieval permissions surface documents users should never see.

Excessive agency

Give agents the narrowest possible tool permissions and require confirmation for writes.

Model supply chain

Downloaded weights and third-party plugins deserve the same scrutiny as npm packages.