Beyond Zero: Enterprise security for the AI era

Google's vision for enterprise security in the age of AI agents. Joseph Valente and Michal Zalewski of Alphabet Security argue that the application-centric zero-trust model (BeyondCorp) is breaking under the strain of autonomous AI agents, exponential data velocity, and machine-speed threats. They propose Beyond Zero, a new paradigm that shrinks the trust boundary from the application level down to individual actions on specific resources.

Key claims:

• Access velocity has outgrown human-speed authorization. AI agents access data at ~10× the rate of humans, reasoning across unstructured datasets. Legacy infrastructure can't authorize tens of millions of concurrent machine-driven actions per second.
• The trust boundary must move. Instead of granting access to an application or tool, Beyond Zero evaluates every action on every resource in realtime, blending static baseline policies (the "floor") with a dynamic AI reasoning engine (the "ceiling").
• Four-component architecture. Autonomous governance precomputes an enterprise world model (who, what, how); event intake streams signals from servers, clients, and agents; a hierarchical reasoning engine makes allow/deny/challenge decisions at access time or via slower deeper inference; challenge infrastructure issues granular interruptions—justifications, security-key touches, biometric checks, approvals—or durable containments that revoke access.
• Challenges replace blunt denials. When risk is ambiguous, the system asks for context rather than blocking outright, reducing disruption for legitimate users while stopping attackers who can't satisfy tailored hard challenges.
• Attackers have weaponized AI too. LLMs rewrite malicious code on demand; agents exploit ambient authority (inheriting overprovisioned human permissions). Beyond Zero counters with intent-based defense and realtime investigation rather than after-the-fact SecOps review.

The article ends with a call for industry standards: open architectures for agent introspection, standardized agentic-identity and request-annotation frameworks, and pluggable policy-evaluation points in SaaS products. It frames the transition as a strategic necessity—security as an immune system that continuously adapts, not a static perimeter.