Definition
When an AI agent is given—or accumulates—more permissions, tools, and capabilities than it actually needs to do its job. If the agent is then tricked or hijacked, those excess permissions become the attacker's leverage, allowing actions far beyond what was intended—like sending emails, modifying files, or escalating to admin access.
Why it matters
Most enterprise AI agent deployments inherit broad developer or service-account credentials by default, creating a massive blast radius if the agent is compromised. Applying the principle of least privilege—giving agents only the minimum permissions needed—is the primary defense.