Definition
An attack where an AI agent gains access to elevated privileges or capabilities beyond what it was authorized to use. This can happen through configuration errors, weak permission checks, or vulnerabilities in the agent framework, allowing the agent to access sensitive data or perform restricted operations.
Why it matters
An agent designed to read customer names from a database but that escalates to admin privileges can now delete records, exfiltrate entire databases, or modify other systems. Privilege escalation turns a limited exposure into a full system compromise.