Not All Agents Are Equally Dangerous: Authority Blast Radius in Watchlight AI Beacon

Every CISO running AI agents in production eventually asks the same question: how worried should I be about this agent? Authority Blast Radius is the answer Watchlight AI Beacon gives, in a form a security team, a compliance officer, and an operator can all use.

Bottom line for CISOs. Not every AI agent in your environment deserves the same controls. A read-only research agent and a payment-processing agent both deserve governance, but they don't deserve the same governance. Authority Blast Radius (ABR) classifies every agent on a four-tier scale your security team understands, and the tier drives which controls Watchlight AI Beacon applies. Higher blast radius, stricter posture. Proportional governance, not blanket governance.

The question every operator eventually asks

Your enterprise has fifty AI agents in production. Maybe two hundred. Maybe more once you count the ones business units stood up without telling security. Each one holds credentials, calls tools, and operates with some degree of autonomy.

Your CISO walks into your office on a Monday morning with a single question: which of these agents should I worry about most?

The honest answer most security teams have today is "I don't know." The audit log shows what each agent did yesterday. It does not show what each agent could do tomorrow if an attacker took over its prompt or its credentials. Without that picture, prioritization is impossible. So is proportional control. So is a defensible conversation with the board.

Why "every agent is the same" is a broken model

A research-summary agent and a payment-processing agent don't deserve the same controls.

The first impulse with AI agent governance is to write one policy that covers everything. Block these tools. Require approval for those. Log all the things. The trouble is that the cost of strict controls is the same for low-risk and high-risk agents, but the value isn't. Force a read-only research agent through human approval for every action and you've made it useless. Skip human approval on a payment-processing agent and you've made it a liability.

What every operator actually needs is a way to classify the agents in front of them, then apply controls proportional to the classification. The same governance plane handles every agent. The intensity of the controls scales with how much damage each one could cause if it went wrong.

What Authority Blast Radius actually measures

ABR quantifies the maximum operational damage an AI agent could cause if compromised. It is derived from twelve declared inputs across three dimensions.

Authority Breadth: what the agent can do

Write-capable tool count. How many systems can this agent actually modify, not just read?
Maximum delegation depth. How far down the delegation chain can it pass authority?
Cross-tenant access. Can this agent reach data outside the tenant or customer it was deployed for?
Administrative authority. Does this agent hold any admin-grade credentials?

Autonomy: how independently the agent operates

Self-spawn capability. Can this agent create new agents on its own?
Plan-modification capability. Can it alter its declared plan mid-execution?
Trust level. What level of trust has the operator assigned to it?
Human gate requirement. Are humans required in the loop for sensitive actions, or not?

Reach: what kinds of damage are reachable

External network access. Can the agent reach the public internet directly?
Financial systems access. Can it touch payments, ledgers, or trading?
PII access. Can it read or modify personal data?
Destructive operations. Can it delete, revoke, or otherwise produce irreversible side effects?

Every input is a declared property of the agent. Watchlight AI Beacon doesn't guess; the classification is derived from facts the operator or registration process states explicitly. That makes the tier reproducible, auditable, and defensible.

Four tiers your security team already understands

Tier	Posture
Constrained	Read-only, sandboxed, no external network. Standard authorization only.
Standard	Single-tenant read or write. Per-tool authorization. Periodic audit.
Elevated	Cross-tenant access or destructive-tool grants. Per-action authorization. SOC review of decisions.
Custodial	Financial, PII, or administrative authority. Default-deny plus mandatory human-approval gate for destructive actions. Continuous drift monitoring. Temporal policy enforcement.

The tier label is what shows up in the dashboard, in the audit log, and in the conversation between the security team and the business owner. A line manager doesn't need to read a continuous risk number; they need to know whether the agent their team deployed is Standard or Custodial. The tier gives them that.

The tier also drives what controls Watchlight AI Beacon applies. A Constrained agent gets light-touch authorization. A Custodial agent gets the full stack: default-deny with human approval for destructive actions, continuous behavioral drift monitoring, temporal policies that span across actions, and full execution lineage. The operator doesn't hand-tune which controls apply where. ABR does.

How ABR shows up in the day-to-day

When an agent is registered with Watchlight AI Beacon, the tier shows up immediately in the registry and in the operator console. From that point on, three things happen continuously:

The tier drives policy selection. Beacon's authorization layer looks at the agent's tier when evaluating each action. A Custodial-tier agent attempting a destructive action triggers default-deny plus human approval, even if the underlying policy alone would have allowed it. The tier adds a strictness floor.

The tier prioritizes operator attention. The operator console highlights Custodial-tier agents at the top of dashboards, drift timelines, and audit views. When operator attention is scarce, ABR makes sure it goes to the agents where the cost of inattention is highest.

The tier is recomputable. When an agent's declared capabilities change (a new tool added, a delegation depth extended, a category of resource access granted), the operator presses Recompute. Watchlight AI Beacon re-evaluates the tier. Every recompute is captured in the tamper-evident audit log, so the history of how an agent's risk profile evolved is reconstructable.

How ABR fits the broader runtime governance posture

Watchlight AI Beacon governs every agent. ABR makes that governance proportional.

Watchlight AI Beacon enforces authorization on every action, regardless of tier. Every action produces a structured audit event. Every drift signal feeds the observability layer. None of that depends on the tier label.

What the tier label does is shape the intensity of the controls applied. The same authorization layer enforces stricter policy when the actor is Custodial. The same drift detector wakes the SOC team at a lower threshold when the actor is Custodial. The same execution-lineage system surfaces faster review queues when the actor is Custodial. ABR is the dial that turns the controls up or down per agent, against the same control plane.

ABR maps cleanly to Principle 3 of the 12 Non-Negotiable Principles for Agent Runtime Governance: Authority Is Explicit, Scoped, and Time-Bound. ABR makes authority explicit by deriving the tier from declared capabilities. It makes authority scoped by mapping tiers to differentiated control postures. It makes authority time-bound by recomputing the tier whenever declared capabilities change.

What ABR enables for the security organization

Proportional governance. Stop treating every agent as equally risky. The Constrained-tier agents your research team stood up over the weekend don't need the same controls as your AP automation agent. ABR makes the difference explicit and applies the appropriate posture automatically.

Prioritized attention. When a drift signal fires, the operator looks at the Custodial tier first. When the SOC reviews policy denials, Custodial sits at the top of the queue. When the compliance officer asks "which agents are most worth a deep review?", the answer is a sortable list, not a guess.

Defensible conversation. When the audit committee, the board, or a regulator asks why an agent does or doesn't require human approval for a particular action, the answer is "because it's Custodial-tier under our governance framework, here are the twelve declared inputs that landed it there, and here is the historical record of every recompute." That is a conversation a CISO can have on record.

Closing

Watchlight AI Beacon governs every agent. Authority Blast Radius makes that governance proportional to the damage each one could do.

Curious where your existing agent fleet would land on the four-tier scale? The Watchlight AI Design Partner Program runs a structured ABR classification exercise against your existing agents and produces a private report mapping each one to its tier and the controls it would activate. One to two weeks. The report is yours regardless of outcome.