Amazon Bedrock AgentCore is an AWS service for running AI agents, with runtime infrastructure that includes Firecracker microVMs and agent-related resources such as container images, memories, and permissions.1 A security report described a weakness in which an exposed agent could be prompted to request a credential endpoint from the Instance Metadata Service (IMDS), turning prompt input into a server-side request forgery-style path to temporary cloud credentials.1
The reported weakness involved AgentCore initially using IMDSv1 alongside insufficient network isolation in its microVM setup.1 The researchers said the resulting credentials could be used to enumerate other AgentCore agents in the same AWS region, retrieve agent container images from Amazon ECR, access memory resources and potentially user conversations, modify agent memories, and retrieve secrets from AWS Secrets Manager.1
Zenity Labs reported the issue to AWS in late 2025 and separately raised concern that AgentCore permissions were scoped to all AgentCore resources in a region rather than an individual agent.1 AWS subsequently moved AgentCore to IMDSv2 and addressed the permission issue, although Amazon disputed describing the matter as a vulnerability.1
The incident illustrates that an agent’s exposure to cloud metadata endpoints can combine with broad identity permissions to turn a prompt-manipulation path into access to infrastructure and data.1 Mitigations described in the report include blocking agent access to credential metadata endpoints, using IMDSv2, and constraining permissions to the narrowest necessary resource scope.1
Use Cases
An AWS reference architecture uses a Strands agent running on AgentCore to answer natural-language questions about antimicrobial-resistance genes detected in public bacterial sequencing data. The agent invokes Athena-backed tools to return grounded, data-backed answers, with Bedrock Guardrails in the conversational layer.2
Related Pages
- Amazon Bedrock — Amazon Bedrock provides the broader managed foundation-model and agent platform that Amazon Bedrock AgentCore extends with agent runtime capabilities.
- AI Agent Security — AgentCore’s runtime isolation and permission safeguards are concrete measures for constraining agent actions and reducing security risks.
- Enterprise AI Agent Architecture — AgentCore supplies runtime capabilities that fit into enterprise agent architectures alongside tools, memory, governance, and observability.
- Docker Cloud Sandboxes — Both AgentCore and Docker Cloud Sandboxes address isolation for AI-agent workloads, though they use different platforms and approaches.
- Strands Box — Strands Box and AgentCore both relate to running agents with isolation and controls that constrain potentially risky actions.