Verified, then logged.
Sun 22 Nov 2026 · 00:21 KST / 15:21 UTC한국어 | ENSearch /

Breaking02 / IT

Amazon Quick lets a separate model approve its agent actions

Seen from behind an office worker at a desk, a laptop screen shows a task list paused at an approval dialog with two blank buttons. AI-generated image

Amazon Web Services (AWS) described Smart Approvals, the agent approval feature in Amazon Quick, on 5 Nov 2026. The feature is on by default, so the company advised administrators to check which actions still go to a person.

According to AWS, only actions whose user setting is Ask each time are reviewed. Actions that another control sends to a person, or that a policy denies, are excluded. A separate model, called the classifier, compares each action with the user's request. It then approves the action automatically or returns it to the user.

Compatibility. According to AWS, profiles with deny-by-default enabled must explicitly allow Smart Approvals. A user's decision to turn the feature off applies across conversations and scheduled runs. No account-level setting can turn it back on for that user.

AWS wrote that each automatic approval is single-use and covers only the action and parameters reviewed. The comparison with the user's request also adds a check against indirect prompt injection, in which instructions in retrieved content try to redirect the agent. Even so, the company cautioned that "Model review can still make mistakes".

Data and audit. AWS said content reviewed by the classifier is not used to train models. Permission updates tied to approvals are recorded in AWS CloudTrail, an audit logging service. That record does not prove an action ran. CloudTrail Event history keeps events for 90 days, so longer investigations need a separately configured trail.

Citing research on security warnings, AWS said repeated confirmation requests can reduce attention. It recommended requiring human approval through connector action permissions for actions that move sensitive data.

The AWS post does not name the model used as the classifier or give its error rate. Before wider rollout, the company recommends that a pilot team test high-risk cases, such as sending content outside the organisation.