Risk scoring

Agent Control > Risk lists every registered ability with its score, level, reasons, whether it reads or writes, whether it has an external effect, its owner plugin, any Ability Auditor findings and the default decision. Create rule starts a rule for that ability.

How the score is built#

The score is deterministic: no AI is involved.

Signal	Points
Marked read-only, with no write verb in its name	low level
Marked destructive in its annotations	+40
Write verb in its name (create, update, edit, set, publish, send, import)	+20
Destructive verb in its name (delete, remove, trash, refund, cancel, reset, purge, drop, deactivate, uninstall)	+40
External-effect verb in its name (send, email, mail, notify, refund, charge, pay, capture, install, activate, upgrade, upload, publish)	+40
Sensitive area (option, setting, user, role, plugin, theme, order, refund, payment, price, coupon, stock)	+20
No annotations at all	+10
Exposed publicly on MCP or REST	+5
Ability Auditor finding on this ability (critical / high)	+20 / +10

Level	Score	Default decision
Low	0 to 19	Allow
Medium	20 to 39	Allow (with snapshot and undo)
High	40 to 59	Hold for approval
Critical	60 and over	Hold for approval

Ability Auditor#

If Ability Auditor is active, Ability Gate reads its latest stored audit and adds its critical and high findings to the score. It only reads. Ability Auditor is not required.