CyberStrikeAI used in 16-day intrusion across four AWS accounts with eight stolen identities and 56 userData reads in 141 ms
Matéria consolidada. Reúne 2 matérias que o Tifon tinha publicado separadamente sobre a mesma história. A IA reescreveu o relatório com o que as 2 fontes trouxeram em 16/11/2026 12:06; os links antigos trazem para cá.
Fontes: securityonline.info ↗ elastic.co ↗
Between September 30 and October 16, 2026, an actor ran the open-source AI offensive framework CyberStrikeAI against four AWS accounts, using stolen credentials from eight identities: five IAM service users and three EC2 instance roles (securityonline.info). Elastic Security Labs reported the activity (elastic.co). The source IP 43.240.29.11 made 56 EC2 user data retrieval calls in 141 milliseconds, then went silent for 14 days before returning on October 16 (elastic.co, securityonline.info). Elastic's detection rules raised 78 low or medium severity alerts that looked routine alone and only made sense once correlated (elastic.co). Elastic has not linked the activity to a named group. It rates the tool attribution as high confidence but cannot prove an LLM agent made the decisions, and it does not know how the credentials were obtained (securityonline.info). This is reported as the first observed use of CyberStrikeAI against AWS (securityonline.info).
Detalhes técnicos
- Phase 1 (Sep 30): GetCallerIdentity calls against four service accounts within 51 minutes (securityonline.info).
- Phase 2 (Sep 30): a compromised EC2 role called Secrets Manager GetSecretValue, then made 56 DescribeInstanceAttribute calls in 141 ms to pull userData (securityonline.info).
- Phase 3 (Oct 1-2): the actor revalidated keys and pivoted to a third EC2 role. That role decrypted an SSM SecureString parameter and issued SSM SendCommand within 29 seconds (securityonline.info).
- Phase 4 (Oct 16): the actor returned, revalidated credentials and accessed a fourth account for the first time (securityonline.info).
- Instance role session names were EC2 instance IDs. Elastic concluded the credentials were stolen from running VMs, likely via SSRF or host compromise (securityonline.info). Elastic describes it as likely instance profile compromise (elastic.co).
- The 78 alerts came from six rules, and 65 fired on day one. The activity was bursty and sporadic, with machine-speed bursts that defeated fixed correlation windows (securityonline.info, elastic.co).
- Infrastructure: FOFA data from Sep 29 showed 43.240.29.11 serving the default CyberStrikeAI panel on port 8080 in AS133199 (SonderCloud, Hong Kong). Neighbouring IPs 43.240.29.28 and 43.240.29.33 presented the same self-signed certificate, serial 3356730520974368630 (securityonline.info). Elastic links the infrastructure to three IPs hosting CyberStrikeAI web interfaces (elastic.co).
Remediação
- Contain the identities: deactivate access keys with iam:UpdateAccessKey, revoke temporary tokens with an aws:TokenIssueTime deny policy, and quarantine the user. Elastic demonstrated this as a Lambda workflow (securityonline.info).
- Rotate credentials and replace compromised EC2 instances rather than only rotating keys (elastic.co, securityonline.info).
- Review secrets exposed in EC2 user data and move them into Secrets Manager (elastic.co, securityonline.info).
- Restrict ec2:DescribeInstanceAttribute with IAM or SCPs (securityonline.info).
- Use behavioral detection and correlate alerts by source IP. Hunt in CloudTrail for more than 20 userData reads in 10 minutes, multi-account GetCallerIdentity calls from one external IP, and the three IPs above (elastic.co, securityonline.info).
Relatos divergentes
- Credential theft vector: elastic.co says the credentials were likely obtained via instance profile compromise, while securityonline.info says Elastic concluded they were stolen from running VMs, likely via SSRF or host compromise. Both describe a hypothesis, and Elastic states it does not know how the credentials were obtained. The securityonline.info piece is a later secondary report of Elastic's research, so elastic.co is the primary source.
Linha do tempo
-
CyberStrikeAI used in 16-day AWS intrusion: 8 stolen identities, 56 EC2 userData scripts read in 141 ms
ler a análise desta fonte ocultar
Between September 30 and October 16, 2026, an actor ran the open-source AI offensive framework CyberStrikeAI against four AWS accounts. It used eight stolen identities: five IAM service users and three EC2 instance roles. Elastic Security Labs reported the activity. The source IP 43.240.29.11 read 56 EC2 userData scripts in 141 ms. It then went dormant for 14 days, which defeated fixed correlation windows. Elastic has not linked the activity to a named group. It says the tool attribution is high confidence, but it cannot prove an LLM agent made the decisions. It also does not know how the credentials were obtained. This is reported as the first observed use of CyberStrikeAI against AWS. The facts come from a single news article, and no web research was done.
Technical Details
- Phase 1 (Sep 30): GetCallerIdentity calls against four service accounts within 51 minutes.
- Phase 2 (Sep 30): a compromised EC2 role used Secrets Manager GetSecretValue, then made 56 DescribeInstanceAttribute calls in 141 ms to pull userData.
- Phase 3 (Oct 1-2): the actor revalidated keys and pivoted to a third EC2 role. That role decrypted an SSM SecureString parameter and issued SSM SendCommand within 29 seconds.
- Phase 4 (Oct 16): the actor returned, revalidated credentials and accessed a fourth account for the first time.
- The instance role session names were EC2 instance IDs. Elastic concluded the credentials were stolen from running VMs, likely via SSRF or host compromise.
- The actor triggered 78 alerts across six rules, all low or medium severity, and 65 of them fired on day one. The alerts only made sense once correlated by source IP.
- FOFA data from Sep 29 showed 43.240.29.11 serving the default CyberStrikeAI panel on port 8080 in AS133199 (SonderCloud, Hong Kong).
- Neighbouring IPs 43.240.29.28 and 43.240.29.33 presented the same self-signed certificate, serial 3356730520974368630.
Remediation
- Contain the identities: deactivate access keys with iam:UpdateAccessKey, revoke temporary tokens with an aws:TokenIssueTime deny policy, and quarantine the user. Elastic demonstrated this as a Lambda workflow.
- Replace compromised EC2 instances rather than only rotating keys.
- Move secrets out of userData into Secrets Manager.
- Restrict ec2:DescribeInstanceAttribute with IAM or SCPs.
- Hunt in CloudTrail for more than 20 userData reads in 10 minutes, multi-account GetCallerIdentity calls from one external IP, and the three IPs above.
-
CyberStrikeAI compromises four AWS accounts over 16 days using stolen credentials and rapid API calls
ler a análise desta fonte ocultar
Between September 30 and October 16, 2026, a CyberStrikeAI server exploited stolen credentials from eight AWS identities across four accounts, executing a series of automated API calls. Notably, it made 56 EC2 user data retrieval requests in 141 milliseconds, followed by intermittent activity including secrets harvesting, SSM parameter decryption, and remote command execution. Elastic detection rules generated 78 alerts of medium or low severity, which individually appeared routine but correlated to a sustained intrusion lasting days with a 14-day silent gap. The attacker likely used compromised EC2 instance role credentials obtained via instance profile compromise, with infrastructure linked to three IPs hosting CyberStrikeAI web interfaces. Elastic Security Labs recommends behavioral detection, credential rotation, and review of secrets exposed in EC2 user data. The case underscores challenges in alert correlation for AI-assisted attacks that exhibit bursty and sporadic behaviors at machine speed.