Google’s red-team organization says the arrival of autonomous agents changes the tempo and shape of security testing. Agents can chain discovery, credential use, tool calls, and lateral movement faster than human operators.
The important shift
Traditional red teams often test a defined environment during a bounded engagement. Agentic testing can continuously explore combinations of permissions and actions. That increases coverage, but also raises containment and audit requirements for the defender’s own tools.
What teams should examine
Authorization should be evaluated at each action, not only at login. Agent identity, delegated authority, tool scopes, memory, secrets, approval boundaries, and revocation need separate controls. A model’s refusal behavior is not a substitute for infrastructure enforcement.