Securing AI Agents with Temporal Policies in Amazon Bedrock AgentCore
AWS ML Blog · rss · 2026-08-07
Limitations of Traditional Access Controls
Before AI agents, access controls typically treated each action as an independent event. However, since agents dynamically decide at runtime which tools to call and with what arguments, stateless controls are insufficient. A tool call that seems safe in isolation might be harmful if executed after reading from an untrusted data source.
Mechanism of Temporal Policies
Amazon Bedrock AgentCore introduces Temporal Policies, allowing developers to define stateful rules. These policies run at the AgentCore Gateway perimeter, outside the agent's code, ensuring they cannot be bypassed or manipulated.
- Trajectory-aware: Evaluates authorization based on the current request and prior events (trajectory) within a session.
- Deny by default: Adheres to a deny-by-default and forbid-wins-over-permit principle.
- Unified management: Routes MCP tool calls, agent-to-agent calls, and model inference through the gateway for consistent policy evaluation.
Core Use Cases
- Enforcing output integrity: Ensures arguments passed to the current tool match the output of a prior call, preventing hallucinations or substitutions.
- Enforcing tool-call ordering: Verifies standard operating procedures by requiring tools to be called in the correct sequence.
- Requiring human approval: Blocks destructive actions until explicit human approval is recorded in the trajectory.
- Enforcing data freshness: Requires data lookups to be completed within a specific timeframe to prevent decisions based on stale data.
The post also uses a private banking portfolio agent as an example to demonstrate how temporal policies can prevent erroneous fund transfers, cap cumulative risk exposure, and automatically tighten permissions when operating without human engagement.
More from coding & agent
- Scaling AI Agent Capacity 30x with a Fast Resumable Stateful Sandbox — rakyll · 2026-08-07
- Hermes Agent Adds Local Parsing for All Document Formats — Teknium · 2026-08-07
- Dev shows ultra-fast resumable stateful sandbox for tool calls — rakyll · 2026-08-07
- Asari Co-Inventor Agents Boost Kimi K3 Inference Speed by 32% — yisongyue · 2026-08-07
- AI Red Teaming Tooling: Choosing Between Microsoft Foundry and PyRIT — WirelessLife · 2026-08-07
- Schmidhuber's team proposes Huxley-Gödel Machine for optimal self-improving coding agents — burny_tech · 2026-08-07