Detects credential abuse of AWS Bedrock by identifying the same principal invoking Bedrock model endpoints (InvokeModel, Converse, and streaming variants) from geographically distant source IPs within a short window, i.e., impossible travel. Bedrock runtime calls are logged as CloudTrail management events, so model-invocation logging is not required. The rule aggregates events by identity (issuing role session ARN when present, otherwise aws.cloudtrail.user_identity.arn) within the same account, then computes first/last geolocations, distance, time window, and inferred travel speed. It requires present geo data and at least two events across at least two countries to reduce noise. Trigger conditions use default thresholds: distance >= 500 km and travel speed >= 800 km/h within a 60-minute lookback window (configured with a 10-minute interval in the query). The alert outputs include identity, account, source IP, location details, distance, travel time, and related fields to support investigation. Investigation fields are exposed for rapid triage and correlation with related CloudTrail attributes including user_identity, location, and client information. MITRE ATT&CK mappings include T1528 (Steal Application Access Token) and T1078 (Valid Accounts, with Cloud Accounts subtechnique), while MITRE ATLAS mappings assign AML.T0012 (Valid Accounts), AML.T0091 (Use Alternate Authentication Material), AML.TA0004 (Initial Access), and AML.TA0015 (Lateral Movement). Setup requirements specify that CloudTrail events be ingested via the AWS integration with GeoIP enrichment for source.geo.location and source.geo.country_name. No Bedrock-specific logging is required. False positives may arise from multi-region automation or CI pipelines sharing credentials across distant egress points, or VPN/proxy configurations causing geolocation inconsistencies. Remediation steps include revoking or rotating the affected access key or role session, reviewing bedrock:InvokeModel permissions, and auditing usage for cost controls and quotas. Alert suppression groups by identity with a 4-hour duration to minimize repeated alerts during legitimate multi-region workflows.