Every AI feature is a new attack surface

Short summary

Every AI feature introduces a new attack surface because models take untrusted input and can act on it, making prompt injection analogous to classic injection vulnerabilities. The article recommends treating model input as hostile by default, constraining model access to least-privilege scopes, and logging actions rather than just outputs. These security practices are not exotic — they are the same disciplines that separate production-ready systems from demos.

• •AI features take untrusted input and act on it — the classic shape of a security problem
• •Treat model input as hostile, constrain access to least privilege, log actions not just outputs
• •AI does not get a security exemption for being new and exciting

Generated with AI, which can make mistakes.