How the attack works
AWS Systems Manager manages instances through an agent that is installed on them and that connects out to the service. Two API calls turn an IAM permission into code execution on a host. SendCommand (Run Command) runs a Systems Manager document on one instance, a list of instances or everything that matches a tag, and with the AWS documents AWS-RunShellScript and AWS-RunPowerShellScript the caller supplies the script. StartSession (Session Manager) opens an interactive shell, or a port forwarding tunnel, to a managed node. Neither needs an open inbound port, an SSH key or a password on the host.
AWS is explicit about what this means. The agent runs with root permissions on Linux and SYSTEM permissions on Windows Server, and AWS states that any principal that has been granted permission to send commands to the agent has root or SYSTEM permissions, and that a user who may run any of the predefined documents has administrator access to the node. An attacker who has taken over such a principal therefore owns the instances it can reach: local data, credentials on disk, and the temporary credentials of the instance role from the instance metadata. With a tag-based target, one call reaches a whole fleet.
Systems Manager logs all control plane operations as CloudTrail management events under ssm.amazonaws.com, and AWS names SendCommand and StartSession among them. The events show who started a command or a session and on what. What was typed inside a session is not in CloudTrail: it is only recorded when session logging to Amazon S3 or CloudWatch Logs is configured. AWS publishes no example record for the two calls, so the rules that read a request field say that the field follows the API reference. No incident report was used for this page.
What an attacker needs
- Credentials for a principal allowed ssm:SendCommand or ssm:StartSession on the target instances.
- Instances that are managed nodes, which means that the agent is running and the instance has a role that lets it register with Systems Manager.
Impact
Code execution as root or SYSTEM on every instance the principal can target. From there the attacker reads application secrets and data, takes the temporary credentials of the instance role and acts with its permissions in the account, and moves to whatever the hosts can reach on the network. A port forwarding session gives a tunnel into the VPC.
Detection rules
Systems Manager Command Sent or Session Started
Detects SendCommand and StartSession calls to AWS Systems Manager. The first runs a Systems Manager document on managed nodes, the second opens an interactive shell or a port forwarding tunnel to one node. AWS states that the agent on the node runs with root or SYSTEM permissions and that any principal allowed to send commands to it has those permissions, and it names both calls among the control plane operations that are logged as management events. The rule matches the event names only and covers every document, including custom ones. It is a broad rule: scope it to callers outside your operations and automation roles, or use it as a hunting query. It also matches attempts that failed (errorCode present): check errorCode, because an unauthorized attempt is worth seeing.
title: Systems Manager Command Sent or Session Started
id: e4f8de3c-f772-47ce-8735-2e3ffe70e399
status: experimental
description: >-
Detects SendCommand and StartSession calls to AWS Systems Manager. The
first runs a Systems Manager document on managed nodes, the second opens
an interactive shell or a port forwarding tunnel to one node. AWS states
that the agent on the node runs with root or SYSTEM permissions and that
any principal allowed to send commands to it has those permissions, and
it names both calls among the control plane operations that are logged as
management events. The rule matches the event names only and covers every
document, including custom ones. It is a broad rule: scope it to callers
outside your operations and automation roles, or use it as a hunting
query.
It also matches attempts that failed (errorCode present): check errorCode,
because an unauthorized attempt is worth seeing.
references:
- https://docs.aws.amazon.com/systems-manager/latest/userguide/monitoring-cloudtrail-logs.html
- https://docs.aws.amazon.com/systems-manager/latest/userguide/ssm-agent-restrict-root-level-commands.html
- https://docs.aws.amazon.com/systems-manager/latest/APIReference/API_StartSession.html
author: Forestall
date: 2026-11-18
tags:
- attack.execution
- attack.t1651
logsource:
product: aws
service: cloudtrail
detection:
selection:
eventSource: ssm.amazonaws.com
eventName:
- SendCommand
- StartSession
condition: selection
falsepositives:
- Patch runs, maintenance windows, configuration management and operators who use Session Manager for shell access. Useful only after the automation and operations roles are excluded with a filter_* selection on userIdentity.arn.
level: low
False positives
- Patch runs, maintenance windows, configuration management and operators who use Session Manager for shell access. Useful only after the automation and operations roles are excluded with a filter_* selection on userIdentity.arn.
Session Manager Sessions Started on Many Instances by One Principal
Counts, per calling principal (userIdentity.arn), on how many different targets it started a Session Manager session. Sessions on five or more different managed nodes within 30 minutes is the pattern of someone walking through a fleet, for example to collect the credentials of each instance role. An operator who works on an incident does the same, so the count and the window are starting points to tune. The target is read from requestParameters.target, which follows the Target request parameter of the StartSession API reference: AWS publishes no example record for this event, so confirm the field in your own events. Denied attempts are counted as well. 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 5 different requestParameters.target values within 30 minutes.
title: Session Manager Session Started (Many Targets Base Event)
id: 4e736f6e-178b-4dfe-835a-7f3112c983ef
name: aws_ssm_session_started
status: experimental
description: >-
Base event of the threshold rule in this file. Any StartSession call to
AWS Systems Manager, including attempts that were denied.
logsource:
product: aws
service: cloudtrail
detection:
selection:
eventSource: ssm.amazonaws.com
eventName: StartSession
condition: selection
level: low
---
title: Session Manager Sessions Started on Many Instances by One Principal
id: c78676ef-a12f-4b16-8ad6-9595ebfa8910
status: experimental
description: >-
Counts, per calling principal (userIdentity.arn), on how many different
targets it started a Session Manager session. Sessions on five or more
different managed nodes within 30 minutes is the pattern of someone
walking through a fleet, for example to collect the credentials of each
instance role. An operator who works on an incident does the same, so
the count and the window are starting points to tune. The target is read
from requestParameters.target, which follows the Target request parameter
of the StartSession API reference: AWS publishes no example record for
this event, so confirm the field in your own events. Denied attempts are
counted as well. 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/systems-manager/latest/APIReference/API_StartSession.html
- https://docs.aws.amazon.com/systems-manager/latest/userguide/monitoring-cloudtrail-logs.html
- https://attack.mitre.org/techniques/T1651/
author: Forestall
date: 2026-11-18
tags:
- attack.execution
- attack.t1651
correlation:
type: value_count
rules:
- aws_ssm_session_started
group-by:
- userIdentity.arn
timespan: 30m
condition:
field: requestParameters.target
gte: 5
falsepositives:
- Operators who check many hosts during an incident or a rollout, and tools that open sessions to instances in turn. Exclude those roles or give them a higher count.
- The values of 5 targets and 30 minutes are starting points, not tested limits. Measure what your own operators produce and adjust both.
level: medium
False positives
- Operators who check many hosts during an incident or a rollout, and tools that open sessions to instances in turn. Exclude those roles or give them a higher count.
- The values of 5 targets and 30 minutes are starting points, not tested limits. Measure what your own operators produce and adjust both.
Shell or PowerShell Script Sent to Instances with Systems Manager Run Command
Detects SendCommand calls to AWS Systems Manager that run the AWS documents AWS-RunShellScript or AWS-RunPowerShellScript. With these two documents the caller supplies the script, and AWS states that it runs on the managed node as if the caller were logged on locally, with administrative permissions. AWS names SendCommand among the Systems Manager calls that are logged as management events. The document is read from requestParameters.documentName, which follows the DocumentName request parameter of the Systems Manager API reference: AWS publishes no example record for this event, so confirm the field in your own events. Read the targets of the request (instance IDs or tag-based targets) to see how many hosts were reached. It also matches attempts that failed (errorCode present): check errorCode, because an unauthorized attempt is worth seeing.
title: Shell or PowerShell Script Sent to Instances with Systems Manager Run Command
id: e37e60a3-b19a-48c6-ba4b-619b3d4df808
status: experimental
description: >-
Detects SendCommand calls to AWS Systems Manager that run the AWS
documents AWS-RunShellScript or AWS-RunPowerShellScript. With these two
documents the caller supplies the script, and AWS states that it runs on
the managed node as if the caller were logged on locally, with
administrative permissions. AWS names SendCommand among the Systems Manager
calls that are logged as management events. The document is read from
requestParameters.documentName, which follows the DocumentName request
parameter of the Systems Manager API reference: AWS publishes no example
record for this event, so confirm the field in your own events. Read the
targets of the request (instance IDs or tag-based targets) to see how many
hosts were reached.
It also matches attempts that failed (errorCode present): check errorCode,
because an unauthorized attempt is worth seeing.
references:
- https://docs.aws.amazon.com/systems-manager/latest/userguide/monitoring-cloudtrail-logs.html
- https://docs.aws.amazon.com/systems-manager/latest/APIReference/API_SendCommand.html
- https://docs.aws.amazon.com/systems-manager/latest/userguide/walkthrough-cli.html
- https://docs.aws.amazon.com/systems-manager/latest/userguide/ssm-agent-restrict-root-level-commands.html
- https://attack.mitre.org/techniques/T1651/
author: Forestall
date: 2026-11-18
tags:
- attack.execution
- attack.t1651
logsource:
product: aws
service: cloudtrail
detection:
selection:
eventSource: ssm.amazonaws.com
eventName: SendCommand
requestParameters.documentName:
- AWS-RunShellScript
- AWS-RunPowerShellScript
condition: selection
falsepositives:
- Operators, configuration tools and maintenance windows that run scripts on instances. They use the same roles every time, so add those roles to a filter_* selection on userIdentity.arn and keep alerting on scripts sent by people outside the operations team, by workload roles or by unknown roles.
level: medium
False positives
- Operators, configuration tools and maintenance windows that run scripts on instances. They use the same roles every time, so add those roles to a filter_* selection on userIdentity.arn and keep alerting on scripts sent by people outside the operations team, by workload roles or by unknown roles.
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.
Command or session on an instance followed by use of its role from elsewhere
The usual goal of a command on an instance is the instance role. Credentials taken from the instance metadata are then used from the attacker's own machine, so calls by the role's session arrive from an address that is not the instance.
- A
SendCommandorStartSessionevent fromssm.amazonaws.com. Note the caller inuserIdentity.arnand the target instance. - Within the next hours, CloudTrail events whose
userIdentity.arnis an assumed-role session of the instance role of that instance. - Compare
sourceIPAddressof those events with the addresses of the instance and of your NAT gateways, and alert when it is none of them.
EventsSendCommandStartSession
LimitsIt needs the instance role of each instance and your own address ranges. Calls that the attacker makes from the instance itself come from the expected address and are not found. The time window is a starting point.
First command or session by a principal that has never operated instances
Operators and automation roles use Systems Manager daily. A workload role, an application user or a newly created identity that sends its first command or starts its first session is using a permission it was never meant to exercise.
- Keep a record, over at least 30 days, of the principals (
userIdentity.arn, or the role inuserIdentity.sessionContext.sessionIssuer.arn) that calledSendCommandorStartSession. - Alert when a principal that is not in the record makes either call.
EventsSendCommandStartSession
LimitsNew operators and new automation produce first-time events too. The 30 days are a starting point.
Triage
- Identify the caller (userIdentity.arn), the source IP address and the user agent. Is it an operator, a maintenance window or an automation role that normally manages these instances?
- Read the document name and the targets of the request. A shell script document sent to many instances, or to a tag that covers a fleet, by a principal that has not done this before is the pattern to look at.
- For a session, read the target instance and check whether the caller is on the team that owns it, and whether the session used port forwarding.
- Read the command invocations in Systems Manager and the session logs, where they are kept, for what was actually run.
- Check what the instance role of the target did in CloudTrail after the event, and from which source address.
Response
Once the attack is confirmed:
- Cancel the running command and terminate the open sessions in Systems Manager.
- Contain the principal that sent the command or started the session, and take away its Systems Manager permissions until its use is explained.
- Treat every targeted instance as accessed with root or SYSTEM rights. Isolate it, keep its disks for review, and rebuild it from a known image.
- Revoke the active sessions of the instance roles of the targeted instances, and rotate the secrets stored on or readable from those hosts.
- Review CloudTrail for what the instance roles did after the event, and remove what was created.
Mitigation
- Grant ssm:SendCommand and ssm:StartSession only to the principals that operate the instances, and limit them to instances with specific tags, as AWS recommends for these two actions.
- Limit which documents a principal may run. Do not allow the predefined
AWS-documents to everyone, because AWS states that they give administrator access to the node. - Configure Session Manager to log session activity to Amazon S3 or CloudWatch Logs, and send Run Command output there, so that what was executed can be reviewed.
- Keep instance roles least-privilege, because every command on the host can use the role.
- Remove Systems Manager permissions from workload roles and CI roles that do not operate instances.
Frequently asked questions
Does CloudTrail show which commands were run?
CloudTrail records that `SendCommand` or `StartSession` was called, by whom and against which target. It does not record what was typed in a session. Session content is only kept when Session Manager is configured to log to Amazon S3 or CloudWatch Logs. For Run Command, the status and output of a command can be read back through Systems Manager itself, for example with `ListCommandInvocations`.
Why is permission to send a command the same as being an administrator on the host?
AWS states that the agent runs with root permissions on Linux and SYSTEM permissions on Windows Server, and that any principal that may send commands to the agent has those permissions. AWS also states that a user who may run any predefined document whose name begins with `AWS-` has administrator access to the node.
Which legitimate activity looks the same?
Patch runs, maintenance windows, configuration tools and operators who use Session Manager instead of SSH all produce these events every day. The rules are useful after the automation roles are excluded. What remains is commands and sessions started by people, by workload roles or by principals that have never operated instances.
Has this been used in attacks?
AWS lists Systems Manager command execution in its own threat technique catalog under cloud administration command, with `SendCommand` and `StartSession` as the events to watch. No public incident report was used for this page. It follows what AWS documents about the agent and these two calls.
References
- MITRE ATT&CK T1651: Cloud Administration Command
- Logging AWS Systems Manager API calls with AWS CloudTrail
- Restricting access to root-level commands through SSM Agent
- Walkthrough: Use the AWS CLI with Run Command (predefined documents)
- AWS Systems Manager API Reference: SendCommand
- AWS Systems Manager API Reference: StartSession
- Logging session activity (Session Manager)
- Threat Technique Catalog for AWS: T1651 Cloud Administration Command