Human Approval for High-Risk Actions (Claude Managed Agent)
Overview
The Human Approval for High-Risk Actions (Claude Managed Agent) evaluation checks whether all enabled tools in a Claude managed agent are configured to require human approval before execution. Tools that are explicitly whitelisted are exempt from this requirement. The check covers both built-in agent tools and tools exposed by connected MCP servers.
Metrics
Approval Gate
A binary check: either all enabled, non-whitelisted tools require human approval or they do not.
Motivation
AI agents that can take consequential actions - sending messages, modifying records, executing code - introduce a new category of risk. A successful prompt injection attack, or simply a misunderstanding of user intent, can cause the agent to initiate actions with real consequences before anyone has had the chance to intervene.
Requiring human approval before tool execution is a direct safeguard against this. It does not limit what the agent can do - it ensures that consequential decisions are not made autonomously when the cost of a mistake is high.
Methodology
- Tool discovery: The evaluation retrieves all tools available to the agent via the Anthropic API, including built-in tools and tools exposed by connected MCP servers.
- Permission policy check: For each tool, the permission policy is inspected to determine whether human approval is required before execution. Disabled tools and explicitly whitelisted tools are excluded from the check.
- Scoring: The agent receives a score of
1if all enabled, non-whitelisted tools require approval,0if any do not.
Scoring
Approval Gate
Examples
Clean - all enabled tools require approval
Agent ID: agent_01AAABBBCCCDDDEEEFFFGGG Whitelisted tools: read, glob, grep Tools:
- bash (enabled, requires approval)
- edit (enabled, requires approval)
- write (enabled, requires approval)
- read (enabled, no approval required, whitelisted)
- glob (enabled, no approval required, whitelisted)
- grep (enabled, no approval required, whitelisted)
- web_fetch (disabled)
0 out of 7 tools violate the human approval requirement:
Passing tools:
- bash - enabled, requires approval
- edit - enabled, requires approval
- write - enabled, requires approval
- read - enabled, no approval required, whitelisted
- glob - enabled, no approval required, whitelisted
- grep - enabled, no approval required, whitelisted
- web_fetch - disabled
Flagged - enabled tool without approval and not whitelisted
Agent ID: agent_01AAABBBCCCDDDEEEFFFGGG Whitelisted tools: read, glob, grep Tools:
- bash (enabled, requires approval)
- edit (enabled, no approval required)
- write (enabled, requires approval)
- read (enabled, no approval required, whitelisted)
1 out of 4 tools violates the human approval requirement:
Violations:
- edit - enabled, no approval required, not whitelisted
Passing tools:
- bash - enabled, requires approval
- write - enabled, requires approval
- read - enabled, no approval required, whitelisted