AG-021
Missing Human Approval
high
Page text and source link for your AI assistant. Nothing is sent automatically.
What Foo Guard detects
The agent has high-impact external capabilities while human approval is explicitly disabled.
When this rule triggers
humanApprovalRequired is false and a tool has a consequential capability such as modification, shell execution, database access, or arbitrary HTTP.
Foo Guard evaluates the normalized configuration supplied to the scanner. A finding describes configuration risk; it does not establish that an attack occurred. An absent finding does not prove that an undeclared capability is safe.
How to fix
Add human approval at the point of consequential execution. Approval must constrain what the tool can do, not just appear in an instruction prompt.
Verify the change
Update the actual agent configuration and the underlying permissions or controls, then run the scanner again. Inspect the finding’s evidence path to confirm the intended setting changed. Do not clear a finding by changing a declaration that no longer reflects the deployed agent.