Skip to content
Threat Feed
Subscribe
medium advisory

Detection of Potential Cross-Region Inference Abuse in Amazon Bedrock

Detection of potential cross-region inference abuse in Amazon Bedrock Claude models, which may indicate attempts to bypass regional security controls or perform unauthorized data processing.

This detection focuses on identifying suspicious cross-region inference activity within Amazon Bedrock, specifically targeting the Claude model family. Cross-region inference abuse occurs when a user or principal executes inference tasks in an AWS region that differs from the primary region of the request origin. This behavior may be indicative of attempts to bypass regional governance policies, evade localized monitoring, or exfiltrate data by leveraging infrastructure in regions with different compliance or logging postures. Defenders should monitor for large-volume input token requests that exhibit a geographic mismatch between the source and the inference execution region. This activity is relevant for organizations managing multi-region AI infrastructure and ensures that generative AI usage remains within authorized operational boundaries.

Impact

Successful abuse of this nature could lead to unauthorized use of AI resources, circumvention of organizational data residency requirements, and potential exfiltration of sensitive information processed by large language models. The impact is primarily associated with cloud infrastructure security and compliance violations.

Recommendation

  • Enable Amazon Bedrock model invocation logging to S3 or CloudWatch Logs to capture request and response payloads.
  • Ingest AWS Bedrock logs into your SIEM platform using the Splunk Add-on for AWS.
  • Review detected cross-region inference events to determine if they are authorized multi-region deployments or anomalous attempts to bypass regional security controls.
  • Baseline expected cross-region usage patterns to reduce noise from legitimate development or testing activities.

Immediate actions

Enable and centralize AWS Bedrock model invocation logging for visibility.

Cloud Security 7d

Threat Hunt

Identify all requests where origin_region != inference_region and input_tokens > 2000.

T1599 medium medium confidence hunt now

Data: AWS Bedrock model invocation logs

Mitigations

Implement Service Control Policies (SCPs) to restrict Bedrock access to authorized regions.

medium_term Cloud Security

Cross-region inference