Privilege escalation (AI agents)

Privilege escalation in AI agent systems is any path by which an agent gains capability beyond what its operator intended: assuming elevated roles, borrowing credentials from its environment, chaining tools to reach systems it cannot touch directly, or inheriting broad permissions through delegation.

Agents escalate differently from human attackers. They do not need to exploit a CVE; they can simply use legitimate credentials lying in scope, or ask a more privileged tool to act on their behalf. Tool chaining is the signature move: a low-privilege agent with a "run script" tool effectively holds every permission the script host does. Containment maps each agent's intended reach explicitly and treats reach beyond it as a governed event, with delegation depth and inheritance tracked (see delegation chain).

Related

• Delegation chain
• Excessive agency
• All glossary terms