How the attack works
Amazon Bedrock serves foundation models to whoever holds credentials with Bedrock permissions, and the account owner pays for every token. AWS's Threat Technique Catalog describes the abuse: an actor who controls an AWS identity uses it to send prompts to the hosted models and can sell or share that access, while the account holder is charged. The calls that do this are InvokeModel, InvokeModelWithResponseStream, Converse and ConverseStream.
Around the invocations there is preparation that leaves its own traces. A key that is not allowed to invoke a model produces denied calls while the attacker tries models and Regions. Where a model still needs an agreement or a use case form, CreateFoundationModelAgreement and PutUseCaseForModelAccess request the access. A service quota increase gives room for more traffic. And because model invocation logging records the full request and response of every invocation, an attacker who does not want the prompts to be read removes it with DeleteModelInvocationLoggingConfiguration. GuardDuty has a finding for that last step, which AWS describes as a possible attempt to hide malicious activity.
Bedrock is unusual in that AWS logs the four invocation calls as CloudTrail management events, so they are visible without data events, under the event source bedrock.amazonaws.com. AWS states that all other Bedrock API operations are logged as management events too, and its example record for InvokeModel shows the model in requestParameters.modelId. CloudTrail does not hold the prompts: only model invocation logging does, and AWS documents that it is disabled by default. No incident report was used for this page.
What an attacker needs
- Valid credentials for a principal with Bedrock permissions, often a leaked long-term access key with a broad policy or a long-term Bedrock API key.
- Access to the wanted model in a Region, or the right to request it, and enough quota for the intended traffic.
Impact
Model usage charges on the owner's bill, which grow with every request for as long as the credential works. If the account uses Bedrock itself, the attacker's traffic competes for the same quotas. With invocation logging deleted, the owner cannot tell what was sent to the models in the account's name.
Detection rules
Amazon Bedrock Model Access Agreement Requested or Use Case Submitted
Detects CreateFoundationModelAgreement and PutUseCaseForModelAccess calls to Amazon Bedrock. The first requests a model access agreement for a model, and the second submits the use case form that some model providers ask for before first use. Both are steps that make a model usable in an account, done once by the team that adopts the model. A request for a model that nobody in the organization uses, by a principal with no Bedrock history, can be preparation for abuse of the account. Both operations are in the Bedrock API reference, and AWS states that all Bedrock API operations are logged by CloudTrail. AWS publishes no example record for them, so the rule matches the event names only and its level is low: read the model from the request (the modelId parameter of CreateFoundationModelAgreement). It also matches attempts that failed (errorCode present): check errorCode, because an unauthorized attempt is worth seeing.
title: Amazon Bedrock Model Access Agreement Requested or Use Case Submitted
id: 86178a5f-7848-4ba6-91e2-ee1465fc65f1
status: experimental
description: >-
Detects CreateFoundationModelAgreement and PutUseCaseForModelAccess calls
to Amazon Bedrock. The first requests a model access agreement for a
model, and the second submits the use case form that some model providers
ask for before first use. Both are steps that make a model usable in an
account, done once by the team that adopts the model. A request for a
model that nobody in the organization uses, by a principal with no
Bedrock history, can be preparation for abuse of the account. Both
operations are in the Bedrock API reference, and AWS states that all
Bedrock API operations are logged by CloudTrail. AWS publishes no example
record for them, so the rule matches the event names only and its level
is low: read the model from the request (the modelId parameter of
CreateFoundationModelAgreement).
It also matches attempts that failed (errorCode present): check errorCode,
because an unauthorized attempt is worth seeing.
references:
- https://docs.aws.amazon.com/bedrock/latest/APIReference/API_CreateFoundationModelAgreement.html
- https://docs.aws.amazon.com/bedrock/latest/APIReference/API_PutUseCaseForModelAccess.html
- https://docs.aws.amazon.com/bedrock/latest/userguide/logging-using-cloudtrail.html
- https://attack.mitre.org/techniques/T1496/004/
author: Forestall
date: 2026-11-18
tags:
- attack.impact
- attack.t1496.004
logsource:
product: aws
service: cloudtrail
detection:
selection:
eventSource: bedrock.amazonaws.com
eventName:
- CreateFoundationModelAgreement
- PutUseCaseForModelAccess
condition: selection
falsepositives:
- Teams that start to use a new model, and administrators who enable models for the organization. These requests are rare, so confirm each one with the team that owns Bedrock in that account.
level: low
False positives
- Teams that start to use a new model, and administrators who enable models for the organization. These requests are rare, so confirm each one with the team that owns Bedrock in that account.
Amazon Bedrock Model Invocation Logging Deleted
Detects DeleteModelInvocationLoggingConfiguration calls to Amazon Bedrock. Model invocation logging collects the full request data, response data and metadata of model invocations in a Region, and AWS documents that logs are stored until the logging configuration is deleted. Deleting it ends the only record of what is sent to the models, and GuardDuty reports the same event as DefenseEvasion:IAMUser/BedrockLoggingDisabled, a possible attempt to hide abuse of AI models. AWS states that all Bedrock API operations are logged by CloudTrail. The call has no request parameters, so the rule matches the event name: the Region of the event (awsRegion) says where logging was removed. It also matches attempts that failed (errorCode present): check errorCode, because an unauthorized attempt is worth seeing.
title: Amazon Bedrock Model Invocation Logging Deleted
id: f4d4ae3f-962f-41e3-8240-e78a876dbce7
status: experimental
description: >-
Detects DeleteModelInvocationLoggingConfiguration calls to Amazon Bedrock.
Model invocation logging collects the full request data, response data
and metadata of model invocations in a Region, and AWS documents that
logs are stored until the logging configuration is deleted. Deleting it
ends the only record of what is sent to the models, and GuardDuty reports
the same event as DefenseEvasion:IAMUser/BedrockLoggingDisabled, a
possible attempt to hide abuse of AI models. AWS states that all Bedrock
API operations are logged by CloudTrail. The call has no request
parameters, so the rule matches the event name: the Region of the event
(awsRegion) says where logging was removed.
It also matches attempts that failed (errorCode present): check errorCode,
because an unauthorized attempt is worth seeing.
references:
- https://docs.aws.amazon.com/bedrock/latest/APIReference/API_DeleteModelInvocationLoggingConfiguration.html
- https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html
- https://docs.aws.amazon.com/bedrock/latest/userguide/logging-using-cloudtrail.html
- https://docs.aws.amazon.com/help-panel/guardduty/latest/console/finding-type-defenseevasion-iam-bedrockloggingdisabled.html
- https://attack.mitre.org/techniques/T1496/004/
author: Forestall
date: 2026-11-18
tags:
- attack.impact
- attack.t1496.004
logsource:
product: aws
service: cloudtrail
detection:
selection:
eventSource: bedrock.amazonaws.com
eventName: DeleteModelInvocationLoggingConfiguration
condition: selection
falsepositives:
- Administrators who move invocation logging to another destination by deleting and recreating the configuration, and teardown of test environments. Confirm that logging was configured again shortly after.
level: medium
False positives
- Administrators who move invocation logging to another destination by deleting and recreating the configuration, and teardown of test environments. Confirm that logging was configured again shortly after.
Many Amazon Bedrock Model Invocations Denied for One Principal
Counts, per calling principal (userIdentity.arn), the Bedrock model invocations that were denied. AWS logs InvokeModel, InvokeModelWithResponseStream, Converse and ConverseStream as management events, so denied attempts are visible without data events. Ten or more denied invocations within 10 minutes is the pattern of someone who tries one model and one Region after another with a credential that was never meant to use Bedrock, or that lacks access to the models it asks for. A misconfigured application produces the same burst, so the count and the window are starting points to tune. Read requestParameters.modelId and awsRegion of the denied calls to see what was tried. For role sessions the principal is the session ARN, so one role used with different session names is counted per session.
Fires when one userIdentity.arn has at least 10 matching events within 10 minutes.
title: Amazon Bedrock Model Invocation Denied (Base Event)
id: 29a562b9-3427-4a29-8aa9-c1ef46fd97b5
name: aws_bedrock_model_invocation_denied
status: experimental
description: >-
Base event of the threshold rule in this file. An InvokeModel,
InvokeModelWithResponseStream, Converse or ConverseStream call to Amazon
Bedrock whose errorCode contains AccessDenied.
logsource:
product: aws
service: cloudtrail
detection:
selection:
eventSource: bedrock.amazonaws.com
eventName:
- InvokeModel
- InvokeModelWithResponseStream
- Converse
- ConverseStream
errorCode|contains: AccessDenied
condition: selection
level: low
---
title: Many Amazon Bedrock Model Invocations Denied for One Principal
id: fce787c4-eb08-43a6-b4cc-c8638c9f65f0
status: experimental
description: >-
Counts, per calling principal (userIdentity.arn), the Bedrock model
invocations that were denied. AWS logs InvokeModel,
InvokeModelWithResponseStream, Converse and ConverseStream as management
events, so denied attempts are visible without data events. Ten or more
denied invocations within 10 minutes is the pattern of someone who tries
one model and one Region after another with a credential that was never
meant to use Bedrock, or that lacks access to the models it asks for. A
misconfigured application produces the same burst, so the count and the
window are starting points to tune. Read requestParameters.modelId and
awsRegion of the denied calls to see what was tried. For role sessions
the principal is the session ARN, so one role used with different session
names is counted per session.
references:
- https://docs.aws.amazon.com/bedrock/latest/userguide/logging-using-cloudtrail.html
- https://aws-samples.github.io/threat-technique-catalog-for-aws/Techniques/T1496.A007.html
- https://attack.mitre.org/techniques/T1496/004/
author: Forestall
date: 2026-11-18
tags:
- attack.impact
- attack.t1496.004
correlation:
type: event_count
rules:
- aws_bedrock_model_invocation_denied
group-by:
- userIdentity.arn
timespan: 10m
condition:
gte: 10
falsepositives:
- Applications that were deployed with a role that lacks access to the model they call, which retry until someone fixes the policy. Fix the policy or exclude the role while it is being fixed.
- The values of 10 denied calls and 10 minutes are starting points, not tested limits. Measure what your own applications produce and adjust both.
level: medium
False positives
- Applications that were deployed with a role that lacks access to the model they call, which retry until someone fixes the policy. Fix the policy or exclude the role while it is being fixed.
- The values of 10 denied calls and 10 minutes are starting points, not tested limits. Measure what your own applications produce and adjust both.
Detection logic
Patterns that need more than one event, a baseline or knowledge of your own environment. Each is a hypothesis to build and tune in your SIEM, not a ready rule.
Model access or quota requested by a principal that has never used Bedrock
Teams that build on Bedrock request model access and quota once, through known people. A principal with no Bedrock history that requests access to a model or a higher quota, and starts invoking soon after, is preparing the account for someone else's traffic.
- Keep a record, over at least 30 days, of the principals (
userIdentity.arn) that made calls tobedrock.amazonaws.com. - Collect
CreateFoundationModelAgreementandPutUseCaseForModelAccessevents frombedrock.amazonaws.com, andRequestServiceQuotaIncreaseevents fromservicequotas.amazonaws.comwhose request names a Bedrock quota (the ServiceCode and QuotaCode parameters of the API). - Alert when the caller is not in the record, and look for
InvokeModelorConverseevents by the same caller in the following hours.
EventsCreateFoundationModelAgreementPutUseCaseForModelAccessRequestServiceQuotaIncreaseInvokeModel
LimitsA new project produces the same first-time events. The 30 days are a starting point. The quota request fields follow the Service Quotas API reference, and the service code that stands for Bedrock was not taken from an AWS example, so read it from your own events.
Invocations of a model or in a Region the organization does not use
Applications call a small, stable set of models in the Regions where they run. Successful invocations of another model, or in another Region, by a principal that is allowed to use Bedrock do not look wrong one by one and are only found against that list.
- Keep a list of the model identifiers and Regions your applications use.
- Collect
InvokeModel,InvokeModelWithResponseStream,ConverseandConverseStreamevents frombedrock.amazonaws.comand readrequestParameters.modelIdandawsRegion. - Alert when the model or the Region is not on the list.
EventsInvokeModelInvokeModelWithResponseStreamConverseConverseStream
LimitsIt needs a maintained list, and cross-Region inference can place requests in Regions you did not choose. AWS's example record shows modelId for InvokeModel only, so confirm the field for the other calls.
Triage
- Identify the caller (userIdentity.arn), its access key, the source IP address and the user agent. Is it one of the applications that use Bedrock?
- Read requestParameters.modelId of the invocations and check whether anyone in the organization uses that model and that Region.
- Count the denied calls of the principal, and check whether a model access request or a quota increase followed.
- Check whether model invocation logging is still configured in each Region where the account uses Bedrock.
- Look at what else the credential did, starting with the first call after it appeared at a new source address.
Response
Once the attack is confirmed:
- Deactivate the access key, delete the long-term API key or revoke the role sessions of the principal that made the calls.
- Restore model invocation logging, and review the invocation logs that still exist for what was sent to the models.
- Withdraw model access and quota increases that nobody in the organization requested.
- Find out how the credential was exposed and close that source, then review everything else it did in the account.
- Review the bill for Bedrock in all Regions and report the unauthorized usage to AWS Support.
Mitigation
- Deny Bedrock actions with a service control policy in accounts and Regions that do not use the service, and deny inference on models the organization does not use.
- Grant bedrock:InvokeModel and the related actions only to the roles of the applications that need them, and prefer roles with temporary credentials over long-term access keys and long-term API keys.
- Enable model invocation logging where Bedrock is used, and deny bedrock:DeleteModelInvocationLoggingConfiguration and bedrock:PutModelInvocationLoggingConfiguration to everyone except its administrators.
- Set billing alarms and budgets for Bedrock, so that unexpected usage is noticed within hours.
- Enable GuardDuty, which has findings for disabled Bedrock logging and for suspicious use of Bedrock APIs.
Frequently asked questions
Are model invocations visible in CloudTrail without data events?
Yes. AWS logs `InvokeModel`, `InvokeModelWithResponseStream`, `Converse` and `ConverseStream` as management events, which CloudTrail records by default. Some other runtime operations, such as the bidirectional streaming call and the agent runtime calls, are data events. The management events show the caller and the model, not the prompt.
Where can I see what was sent to the models?
Only in model invocation logging, which collects the full request data, response data and metadata of invocations and delivers them to CloudWatch Logs or Amazon S3. AWS documents that it is disabled by default and that logs are stored until the logging configuration is deleted. That is why deleting the configuration has a rule of its own.
Why is there no rule on InvokeModel itself?
Every application that uses Bedrock produces these events all day, so a rule on a successful invocation alone is noise. The page uses a threshold on denied invocations, which shows a principal trying models it may not use, and single events for the rarer steps around the abuse. Successful abuse by a principal that is allowed to invoke models shows on the bill and in a comparison with normal usage.
Does a quota increase request belong to this attack?
It can. More traffic than the default quotas allow needs an increase, which is requested through Service Quotas with `RequestServiceQuotaIncrease` and is logged by CloudTrail. The request names the service and the quota. An increase for a Bedrock quota that nobody in the organization asked for is worth checking together with the other events on this page.
References
- MITRE ATT&CK T1496.004: Resource Hijacking: Cloud Service Hijacking
- Monitor Amazon Bedrock API calls using CloudTrail
- Monitor model invocation using CloudWatch Logs and Amazon S3
- Amazon Bedrock API Reference: DeleteModelInvocationLoggingConfiguration
- Amazon Bedrock API Reference: CreateFoundationModelAgreement
- Amazon Bedrock API Reference: PutUseCaseForModelAccess
- GuardDuty finding: DefenseEvasion:IAMUser/BedrockLoggingDisabled
- Logging Service Quotas API calls using AWS CloudTrail
- Threat Technique Catalog for AWS: T1496.A007 Bedrock LLM Abuse