Security & Privacy
Cloud Metadata Endpoint (IMDS)
Category
Security & Privacy
Abstraction level
Primitive
Operation level
DeploymentSystem
Use cases
Analyzing SSRF risk in AI agentsSecuring cloud credentialsDesigning sandboxes for agents' network tools
How it works
IMDS responds to HTTP requests from within the instance, returning metadata and credentials. Mitigations include IMDSv2 (a session token in the header, hop-limit restrictions), blocking the link-local address for agent processes, and sandboxing network tools.
Problem solved
How to protect cloud credentials from leaking when an AI agent has network access and could be manipulated into querying instance metadata?
Strengths & limitations
Strengths
✓Convenient, standard mechanism for delivering credentials
✓IMDSv2 significantly reduces abuse
Limitations
✗An unsecured IMDS is a serious credential-leak vector
✗Requires deliberate configuration and isolation of agent tools