> ## Content Index
> Fetch the complete content index at: https://f4n6.co.uk/llms.txt
> Use this file to discover other available public pages before exploring further.

# CyberStrikeAI server abuses stolen AWS creds in 141ms burst
- URL: https://f4n6.co.uk/security-feed/cyberstrikeai-server-abuses-stolen-aws-creds-in-141ms-burst/
- Published: 2026-11-13T23:26:52.000Z
- Updated: 2026-11-13T23:26:52.000Z
- Author: Jeff Davies
- Tags: #security-feed, CyberStrikeAI

## 1\. Executive summary

Between 30 September and 16 October 2026 a server at 43\[.\]240\[.\]29\[.\]11, fingerprinted by Elastic Security Labs as a CyberStrikeAI instance, used credentials belonging to eight AWS identities (five IAM users and three EC2 instance roles) across four accounts, generating 78 CloudTrail alerts. The most distinctive event was 56 `DescribeInstanceAttribute` calls requesting EC2 `userData` in 141 milliseconds from a single assumed-role session, which points to automated tooling rather than a person at a console. On 2 October a third role read an encrypted SSM parameter and issued an SSM `SendCommand` 29 seconds later; the audit log does not record what the command was or whether it succeeded. All 78 alerts were rated medium or low severity and only resolved into one intrusion when correlated by source IP, and the operator returned after a 14-day gap that a fixed correlation window would have missed. No CVE is associated with this item and no exploitation of a named vulnerability is reported; the exposure is stolen cloud credentials, and the risk to EMEA financial services is concentrated in any estate where EC2 instance roles, Secrets Manager and SSM sit in the same trust path as internet-facing workloads. The CyberStrikeAI attribution is Elastic's own and is single-sourced; no MITRE ATT&CK profile exists for the actor, so attribution beyond "a CyberStrikeAI server" is unconfirmed.

## 2\. Regulatory framing

| Article                                                                         | Trigger (the fact in this item)                                                                                                                             | Practical impact                                                                                                                                                                                                              |
| ------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| DORA Art. 19: reporting of major ICT-related incidents to competent authorities | An external IP used credentials from eight IAM identities across four AWS accounts, including an SSM SendCommand whose content and outcome are not recorded | If the same pattern is found in a client estate, the inability to reconstruct what the remote command did bears directly on whether the incident is major and on what can be reported within the DORA timelines               |
| DORA Art. 28: ICT third-party risk — general principles                         | The intrusion ran entirely through AWS-managed services (STS, Secrets Manager, EC2, SSM) under the client's own IAM identities                              | Contractual and control assumptions about the cloud provider's shared-responsibility boundary need checking against what the provider's audit log actually records, since CloudTrail here did not capture the command payload |
| NIS2 Art. 21(2)(d): supply chain security measures                              | Three EC2 instance-role sessions were used from an outside IP, meaning credentials were taken from the instances themselves                                 | Where those instances belong to, or are managed by, a supplier, the incident is a supply-chain event and the supplier's credential handling is in scope                                                                       |

## 3\. Technical analysis & attack chain

Elastic Security Labs published this account on 2026-11-13 and it is the only source for the technical detail below. The victim is not named; accounts and identities are aliased as Accounts A to D. Elastic states it cannot prove an LLM agent was making the decisions and does not know how the attacker obtained the credentials.

Confirmed sequence, in the order Elastic records it:

1. Credential validation. The server called `GetCallerIdentity` on four IAM users across three accounts within 51 minutes. Three of those calls landed in the same second, one in each of three accounts. Two `GetSessionToken` calls on one user requested temporary credentials.
2. Secret and user data harvesting. From `app-role`, an EC2 instance role in Account A, the server read a secret from Secrets Manager at T+3 h 18 min, made one `DescribeInstanceAttribute` user data request at T+6 h 10 min, then fired 56 `DescribeInstanceAttribute` calls requesting the `userData` attribute in 141 milliseconds, about 2.5 ms per call. All 57 calls that day came from the same assumed-role session.
3. Second role and SSM. On 1 October the server revalidated credentials (four `GetCallerIdentity` calls at T+0 s), pulled more secrets at T+1 s, made one more user data request at T+14 s, and read secrets from a non-production environment at T+4 h 10 min with a second pull about three minutes later. On 2 October `container-role` issued `GetParameter` with decryption against an SSM SecureString parameter at T+0 s and an SSM `SendCommand` at T+29 s.
4. Return visit. After 14 quiet days the server revalidated credentials on 16 October, including a user in a fourth account that had not appeared before, at T+1 h 57 min and T+3 h 7 min.

The eight identities and their actions as Elastic tabulates them: `app-role` (EC2 instance role, Account A) ran `GetSecretValue` and `DescribeInstanceAttribute` for 60 alerts between 30 September and 1 October; `logship-user-1` (IAM user, Accounts A and C) ran `GetCallerIdentity` for 5 alerts from 30 September to 16 October; `logship-user-2` (Account B) and `email-svc-user` ran `GetCallerIdentity` and `GetSessionToken` on 30 September; `nonprod-role` ran `GetSecretValue` for 2 alerts on 1 October; `cert-renew-user` and `container-role` ran `GetParameter` and `SendCommand` on 2 October; `named-user` (Account D) ran `GetSessionToken` for 1 alert on 16 October.

The three role sessions were named after EC2 instance IDs, so the credentials came from instance profiles. Elastic's assessment is that using them from an outside IP means they were taken from the instances, for example through SSRF or host compromise. The initial access vector is not established beyond that.

What the harvested data is worth: EC2 user data scripts are the startup scripts attached to an instance and, as Elastic notes, often hold hardcoded API keys, database credentials, bootstrap secrets, SSH keys or certificates. The 56-call burst asked only for the `userData` attribute, and the consistent 2.5 ms spacing is consistent with tooling working through a prepared list of instance IDs. Elastic's guidance is to treat all secrets in EC2 user data as exposed.

The SSM step is the highest-risk activity in the record. Within 29 seconds a third role read an encrypted SSM parameter and issued `SendCommand`. CloudTrail records that a command was sent; it does not record the command or its result. Elastic advises reviewing SSM command history for the window of any unexpected `SendCommand` and checking for abuse of the AWS-managed documents `AWS-RunShellScript` and `AWS-RunPowerShellScript` or custom documents.

Infrastructure. Internet scan data shows 43\[.\]240\[.\]29\[.\]11 serving a CyberStrikeAI web interface on port 8080 with a self-signed certificate, one day before the first alert. A FOFA search for `ip="43.240.29.11" && title=="CyberStrikeAI"` returns two records, one HTTP and one HTTPS, both on port 8080; the HTTPS record was seen 2026-09-29 and the HTTP record 2026-10-06\. The host is in AS133199, SonderCloud Limited, Hong Kong SAR, and responds HTTP/2.0 200 OK with permissive CORS headers allowing POST, OPTIONS, GET, PUT and DELETE. The TLS certificate is self-signed and untrusted, issuer and subject CN `CyberStrikeAI`, ECDSA with SHA-256, valid 2026-09-15 to 2027-09-15\. Two further IPs, 43\[.\]240\[.\]29\[.\]28 and 43\[.\]240\[.\]29\[.\]33, serve the same page on port 8080 with the same self-signed certificate, same hosting, same FOFA fid and same serial number, which ties the three together. As of November none of the three is serving that interface, so the IPs are for retrospective hunting rather than blocking. Port 8080 is the web interface port in the platform's default setup.

Tooling context. CyberStrikeAI is an open-source platform that pairs LLM agents with offensive security tools. Elastic cites Will Thomas at Team Cymru as having counted 21 IPs running it in early 2026, all aimed at Fortinet edge devices; Elastic describes this AWS case as the first it has seen of the same tooling turned on AWS.

Detection coverage and its limits. Six Elastic detection rules fired across the four phases: AWS EC2 User Data Retrieval for EC2 Instance (58 alerts, medium, phases 2 and 3), AWS STS GetCallerIdentity API Called for the First Time (11, phases 1, 3 and 4), First Time Seen AWS Secret Value Accessed in Secrets Manager (4, low, phases 1 and 4), AWS SSM SendCommand Execution by Rare User, and AWS Systems Manager SecureString Parameter Request with Decryption Flag. Elastic reports 65 of the 78 alerts (83%) on day one, then a few targeted actions per day. Every alert was medium or low severity and routine in isolation; the shared source IP is what joined them.

Elastic maps the activity to five ATT&CK techniques: T1078.004 Valid Accounts: Cloud Accounts (initial access), T1087.004 Account Discovery: Cloud Account (discovery), T1552.005 Unsecured Credentials: Cloud Instance Metadata API and T1555.006 Credentials from Password Stores: Cloud Secrets Management Stores (credential access), and T1651 Cloud Administration Command (execution).

Confidence caveats. The CyberStrikeAI attribution rests on Elastic's correlation of scan data (page title, certificate, FOFA fid) with its own AWS telemetry, and is single-sourced; verify before acting on it as an actor identification. Elastic's "AI-assisted" label rests on the tool attribution and the machine-speed timing, and Elastic states plainly that it cannot prove an LLM agent made the decisions. The 14-day gap is consistent with a cool-down, a return after working other targets, or a check that access survived; Elastic states the data does not separate these, and neither does this advisory.

## 4\. Mitigation & containment

P1 within 24 hours

- Hunt CloudTrail for the four behaviours Elastic lists: a burst of `DescribeInstanceAttribute` calls requesting `userData`, many per second, from one role session; first-time `GetCallerIdentity` calls from the same external IP across several accounts or users within the same minute; `GetSessionToken` or `GetSecretValue` from an instance role session whose source IP is not the instance or your NAT; and `GetParameter` with decryption followed within a minute by `SendCommand` from the same role.
- Search historical CloudTrail for 43\[.\]240\[.\]29\[.\]11, 43\[.\]240\[.\]29\[.\]28 and 43\[.\]240\[.\]29\[.\]33\. The three IPs are no longer serving the interface, so this is retrospective hunting, not live blocking.
- For any identity that touched those IPs, rotate credentials and revoke active sessions, including sessions from temporary credentials. For instance roles, rotate by replacing the instance profile or the instance, because stolen role credentials expire but can be reissued while the host stays compromised.
- Review SSM command history for the window of any unexpected `SendCommand` to establish what ran, and check for use of `AWS-RunShellScript`, `AWS-RunPowerShellScript` or custom documents.
- Treat everything an affected role could read as compromised and work outward from the credentials. Treat all secrets in EC2 user data as exposed.

P2 within 72 hours

- Move secrets out of EC2 user data into Secrets Manager or Parameter Store.
- Restrict `ec2:DescribeInstanceAttribute` for `userData` to the roles that need it, using an IAM policy or a service control policy.
- Check non-production roles and secrets for overlap with production. Elastic's point is that non-production environments often have weaker controls and sometimes share secrets with production, so treat access to one as access to the other until shown otherwise.
- Enable GuardDuty where it is not already on, and alert on credentials used from outside expected IP ranges.
- Block the three IPs at the WAF, VPC, network ACL and security group level. Elastic cautions that network blocking does not prevent stolen credentials being used against API calls, so pair it with the behavioural detections.

P3 within 7 days

- Hunt for how the instance credentials were obtained, starting with exposed web applications, SSRF, code repositories and the instance metadata service.
- Deploy the two high-signal detections Elastic published for this behaviour: AWS EC2 Bulk User Data Enumeration, which alerts when one principal and source IP read `userData` from more than 20 distinct instances in 10 minutes, and AWS STS GetCallerIdentity Bulk Account Enumeration, which alerts when three or more distinct non-role identities call `GetCallerIdentity` from the same external IP and ASN in 10 minutes.
- Where Elastic Security 9.4+ is deployed, Elastic Workflows can attach automated response to these alerts. Elastic's proof of concept triggers on the bulk user data rule, iterates the user names in the alert, and through the AWS Lambda connector deactivates the compromised access key with `iam:UpdateAccessKey`, blocks older temporary sessions by adding an inline deny policy via `iam:PutUserPolicy` with an `aws:TokenIssueTime` cutoff, and quarantines the IAM user with an inline deny-all policy. The example workflow is at https://github.com/elastic/workflows.

No fixed release or vendor patch applies to this item; the remediation is credential rotation, IAM scoping and secret relocation. No CVE is associated with it.

## 5\. Indicators of compromise

| type                   | value                                                                                                                                                                                           | confidence                                                                             | source                |
| ---------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------- | --------------------- |
| ipv4                   | 43\[.\]240\[.\]29\[.\]11                                                                                                                                                                        | high (source of all 78 alerts)                                                         | Elastic Security Labs |
| ipv4                   | 43\[.\]240\[.\]29\[.\]28                                                                                                                                                                        | medium (same certificate, hosting and FOFA fid; not in this investigation's telemetry) | Elastic Security Labs |
| ipv4                   | 43\[.\]240\[.\]29\[.\]33                                                                                                                                                                        | medium (same certificate, hosting and FOFA fid; not in this investigation's telemetry) | Elastic Security Labs |
| x509 serial            | 3356730520974368630                                                                                                                                                                             | high (shared by all three IPs)                                                         | Elastic Security Labs |
| x509 sha256 public key | 45:FD:83:A2:38:2C:DD:70:2A:97:1D:E4:66:57:7D:E9:B0:EA:41:48:80:8A:B9:D1:63:95:90:C5:1E:A7:04:94:E3:21:7C:73:9F:EA:D2:29:3F:31:66:87:75:EC:BA:83:23:59:CA:4D:D9:7E:C2:9A:FC:FE:7D:28:5A:31:1C:67 | high (shared by all three IPs)                                                         | Elastic Security Labs |

The three IPs are in AS133199, SonderCloud Limited, Hong Kong SAR. The certificate is self-signed, issuer and subject CN `CyberStrikeAI`, ECDSA with SHA-256, valid 2026-09-15 04:13 UTC to 2027-09-15 05:13 UTC, with subject alternative names `localhost`, 127.0.0.1 and ::1\. The loopback names and addresses are certificate fields, not pivotable indicators. The certificate serial and public key are the durable pivots if the operator moves to new addresses.

```iocs
ipv4  43[.]240[.]29[.]11
ipv4  43[.]240[.]29[.]28
ipv4  43[.]240[.]29[.]33
x509_serial  3356730520974368630
x509_sha256_pubkey  45:FD:83:A2:38:2C:DD:70:2A:97:1D:E4:66:57:7D:E9:B0:EA:41:48:80:8A:B9:D1:63:95:90:C5:1E:A7:04:94:E3:21:7C:73:9F:EA:D2:29:3F:31:66:87:75:EC:BA:83:23:59:CA:4D:D9:7E:C2:9A:FC:FE:7D:28:5A:31:1C:67

```

Behavioural indicators

| behaviour                                                                                                            | where to observe | confidence                                                                  |
| -------------------------------------------------------------------------------------------------------------------- | ---------------- | --------------------------------------------------------------------------- |
| Burst of DescribeInstanceAttribute calls requesting userData, many per second, from one role session                 | CloudTrail       | high                                                                        |
| First-time GetCallerIdentity calls from the same external IP across several accounts or users within the same minute | CloudTrail       | high                                                                        |
| GetSessionToken or GetSecretValue from an instance role session whose source IP is not the instance or your NAT      | CloudTrail       | high                                                                        |
| GetParameter with decryption followed within a minute by SendCommand from the same role                              | CloudTrail       | high                                                                        |
| Role session names matching EC2 instance IDs used from an outside IP                                                 | CloudTrail       | medium (Elastic's inference that credentials were taken from the instances) |

## 6\. Detection

The sources give no file artefacts, so no YARA rule is possible. The behavioural material supports a Sigma rule for the SSM sequence, which is the highest-risk step in the record.

```yaml
title: AWS SSM Parameter Decryption Followed by SendCommand From Same Role
id: 8f3c2a71-4d6e-4b19-9a52-7c1e0d4b6f83
status: experimental
description: Detects an SSM SecureString parameter read with decryption followed within 60 seconds by an SSM SendCommand from the same assumed-role session, the sequence Elastic Security Labs recorded on 2026-10-02.
author: Adverse Trace
date: 2026-11-13
references:

  - https://www.elastic.co/security-labs/threat-command/cyberstrikeai-aws-cloudtrail-intrusion
logsource:
  product: aws
  service: cloudtrail
detection:
  selection_getparameter:
    eventSource: ssm.amazonaws.com
    eventName: GetParameter
    requestParameters.withDecryption: 'true'
  selection_sendcommand:
    eventSource: ssm.amazonaws.com
    eventName: SendCommand
  timeframe: 60s
  condition: selection_getparameter | near selection_sendcommand
fields:

  - userIdentity.arn
  - sourceIPAddress
  - requestParameters.name
falsepositives:

  - Automation that reads a SecureString parameter and immediately runs a document through SSM under the same role
level: high

```

Elastic's own two ES|QL detections are worth deploying as published rather than reimplemented: AWS EC2 Bulk User Data Enumeration (one principal and source IP reading `userData` from more than 20 distinct instances in 10 minutes) and AWS STS GetCallerIdentity Bulk Account Enumeration (three or more distinct non-role identities calling `GetCallerIdentity` from the same external IP and ASN in 10 minutes). Both are described in the Elastic report cited in §7.

## 7\. Sources

- Elastic Security Labs, "56 AWS API calls in 141 milliseconds: Inside an AI-run intrusion", https://www.elastic.co/security-labs/threat-command/cyberstrikeai-aws-cloudtrail-intrusion, 2026-11-13
- Elastic Security Labs, "CyberStrikeAI hit AWS: 56 API calls in 141 milliseconds", https://www.elastic.co/security-labs/threat-command/cyberstrikeai-aws-cloudtrail-intrusion (same report, alternate presentation)
- Team Cymru (Will Thomas), "Tracking CyberStrikeAI Usage", https://www.team-cymru.com/post/tracking-cyberstrikeai-usage, cited by Elastic; not independently reviewed for this advisory
- CyberStrikeAI project, https://github.com/Ed1s0nZ/CyberStrikeAI
- Elastic Workflows, https://github.com/elastic/workflows

## 8\. Adverse Trace position

Adverse Trace assesses this as a moderate-severity cloud credential compromise with a high-severity unresolved step: the 2 October `SendCommand` 29 seconds after an encrypted parameter read is the one action in the record whose content and outcome are unknown, and until SSM command history is reviewed for that window the impact cannot be bounded. There is no CVE, no named vulnerability and no CISA KEV entry, so severity here comes from the credential exposure and the audit gap, not from a scored flaw. The CyberStrikeAI attribution is Elastic's and is single-sourced; no MITRE ATT&CK profile exists for the actor, so we treat attribution beyond "a CyberStrikeAI server" as unconfirmed and will not name an operator. The direct victim is not identified and no source links this intrusion to financial services, so client exposure is indirect: any EMEA financial services estate running EC2 instance roles, Secrets Manager and SSM in one trust path should assume the same pattern is possible and hunt for it. We will monitor for further reporting that identifies the victim, establishes the initial access vector, or recovers the SSM command content, and we will reissue this advisory if any of those appear. Clients who find the four CloudTrail behaviours in their own logs should contact us before rotating credentials, so the session revocation and the forensic capture happen in the right order.

---

[Read the original source →](https://www.elastic.co/security-labs/threat-command/cyberstrikeai-aws-cloudtrail-intrusion?ref=f4n6.co.uk)

*Published via PulseTrace — Adverse Trace threat intelligence.*