SPLHunting query: tune before alertingCloud AI abuse: Amazon Bedrock (LLMjacking)
Bedrock invocation probing: failed model calls across models or regions
LLMjacking operators test stolen AWS keys by sending deliberately invalid InvokeModel requests (Sysdig saw max_tokens_to_sample set to -1): a ValidationException proves the key can reach the model, AccessDenied shows it cannot. This query groups failed Bedrock runtime calls per principal and hour and surfaces principals that fail across several models or regions, or many times.
- Log source
- AWS CloudTrail in Splunk (sourcetype aws:cloudtrail), eventSource bedrock.amazonaws.com
- Techniques
- LLM10 Unbounded Consumption · AML.T0034 Cost Harvesting · AML.T0040 AI Model Inference API Access · AML.T0012 Valid Accounts
- MITRE ATT&CK
- T1078.004Cloud AccountsT1496.004Cloud Service Hijacking
- Reviewed
- 2026-11-10
Query · SPL
index=aws sourcetype=aws:cloudtrail eventSource="bedrock.amazonaws.com"
eventName IN ("InvokeModel", "InvokeModelWithResponseStream", "Converse", "ConverseStream")
errorCode IN ("ValidationException", "AccessDenied", "AccessDeniedException")
| eval principal=coalesce('userIdentity.arn', 'userIdentity.accessKeyId')
| bin _time span=1h
| stats count AS failed_calls dc(awsRegion) AS regions dc(requestParameters.modelId) AS models
values(errorCode) AS errors values(userAgent) AS user_agents values(sourceIPAddress) AS source_ips
BY _time principal
| where regions >= 2 OR models >= 3 OR failed_calls >= 10
| sort - failed_calls Other platforms
Microsoft Sentinel / Defender (KQL)
Microsoft Sentinel / Log Analytics: AWSCloudTrail (Amazon Web Services S3 connector)
AWSCloudTrail
| where TimeGenerated > ago(1d)
| where EventSource == "bedrock.amazonaws.com"
| where EventName in (dynamic(["InvokeModel", "InvokeModelWithResponseStream", "Converse", "ConverseStream"]))
| where ErrorCode in ("ValidationException", "AccessDenied", "AccessDeniedException")
| extend Principal = coalesce(UserIdentityArn, UserIdentityAccessKeyId),
ModelId = tostring(parse_json(RequestParameters).modelId)
| summarize FailedCalls = count(), Regions = dcount(AWSRegion), Models = dcount(ModelId),
Errors = make_set(ErrorCode), UserAgents = make_set(UserAgent, 20), SourceIPs = make_set(SourceIpAddress, 20)
by Principal, HourBucket = bin(TimeGenerated, 1h)
| where Regions >= 2 or Models >= 3 or FailedCalls >= 10
| order by FailedCalls desc Expected false positives
- Developers trying a new model or request schema and hitting validation errors
- SDK or model-version upgrades that change request parameters
- Integration tests that deliberately exercise error paths
Sources
- Sysdig: LLMjacking — stolen cloud credentials used in new AI attack (May 2024)
- Permiso: Exploiting hosted models — LLMjacking on Amazon Bedrock (Oct 2024)
- AWS docs: Monitor Amazon Bedrock API calls using CloudTrail
Reviewed against vendor documentation, not tested in your environment. Defensive use only.