Google ADK

Agent framework · Google · 4/7 verified · Last checked 2026-10-16

Tool confirmation can pause a call for a yes or no, and before-tool callbacks can block calls that break policy. Sandboxing and network perimeters are left to you.

Overall: C (1.50 avg, 4/7 controls verified)

Lock this down first

Add a before_tool_callback that validates arguments, and use a sandboxed code executor rather than local execution.

Scorecard

Criterion	Rating	Notes
Permission granularity	2/3	Before-tool callbacks and in-tool policy checks can block calls.
Human approval gates	2/3	require_confirmation pauses a tool for approval. The feature is marked experimental.
Sandboxing	Unverified	Not confirmed in the docs we checked.
Secrets handling	Unverified	Not confirmed in the docs we checked.
Audit logs	2/3	Python logging plus OpenTelemetry GenAI logs. Logging plugins capture tool calls.
Kill switch / interrupt	Unverified	Not confirmed in the docs we checked.
Network egress control	0/3	Docs advise confining agents inside VPC Service Controls perimeters, a Google Cloud control you set up yourself. ADK itself provides none.

Unverified: Sandboxing, Secrets handling, Kill switch / interrupt. That means we could not confirm it in public docs, not that it is missing.

Sources

Vendor? Request a correction. We review corrections within 7 days.