이미지: AWS ML Blog
Summary
- AWS added a "temporal policies" feature to Amazon Bedrock AgentCore
- It determines whether to approve a current request based on the agent's prior tool-call history
- It runs at the gateway boundary, so it cannot be bypassed by the agent's own code
- 발표 매체
- AWS ML Blog
- 작성자
- Sean Eichenberger
- 발행일
- 2026-08-06
- 적용 대상
- Amazon Bedrock AgentCore Gateway
- 주요 기능
- 워크플로 순서 강제, 데이터 조작 방지, 재무 노출 한도, 고액 작업 승인 요구
AWS announced via its ML blog that it has added "temporal policies" to Amazon Bedrock AgentCore. While existing policies were stateless, evaluating each request independently, the new feature also takes into account an agent's session history — the sequence of its previous tool calls — when deciding whether to approve a current request.
Why it's needed
The blog explained that because AI agents decide at runtime which tools to call, with what arguments, and in what order, an individual call may look safe on its own but become dangerous when combined with prior context. Examples cited include an agent transferring funds using an incorrect account number after having looked up the correct one, continuing to execute repeated trades even after exceeding a cumulative loss limit, and both approving and rejecting the same insurance claim within seconds of each other. AWS noted that such cases cannot be caught by inspection at the level of individual calls.
What's changing
Temporal policies run at the AgentCore Gateway boundary, outside the agent's own code, so they cannot be bypassed through prompt manipulation or code bugs, AWS explained. Use cases introduced include "output integrity enforcement," which prevents value tampering by only allowing arguments that exactly match the results of previous tool calls; "tool call sequence enforcement" to ensure compliance with specific procedures; a feature requiring a recorded, explicit human approval before sensitive operations; and "data freshness enforcement," which ensures only the most recently retrieved data within a set time window is used. AWS added that the feature can also be used to automatically narrow permissions for agents that operate without human intervention.
