AI-Assisted Hackers Compromise AWS Cloud in 72 Hours Using Stolen Credentials

An AI-assisted threat actor has demonstrated how quickly a modern AWS environment can be compromised when valid credentials, weak identity controls, and exposed secrets intersect. In about 72 hours, they achieved broad cloud control using familiar techniques executed at an unprecedented speed.

This incident, investigated by a cloud security team, reveals that attackers no longer require zero-day vulnerabilities or custom malware. Instead, they can use artificial intelligence to compress reconnaissance, scripting, and lateral movement into tightly orchestrated attack waves.

Compromise AWS Cloud in 72 Hours

The intrusion began with a vulnerability in an internet-facing application that exposed an AWS access key, giving the attacker an initial foothold in one of the organization’s cloud accounts. From this starting point, the attacker employed AI-assisted tools to quickly enumerate services, identities, permissions, and deployment workflows.

This allowed them to identify misconfigurations and overly permissive roles that could be exploited for deeper access. Each newly discovered credential or token became a pivot point for further exploitation: the attacker validated each one, mapped its effective permissions, and then repeated discovery, secrets harvesting, and persistence steps within the newly expanded scope.

According to Sygnia, the campaign unfolded across multiple layers of the AWS environment, including ECS and EC2 workloads, S3 storage, RDS databases, CI/CD runners, and source control platforms such as GitHub and Bitbucket.

Secrets were systematically collected from environment variables, plaintext data in S3, application databases, AWS Secrets Manager, and Systems Manager Parameter Store. This collection provided the necessary connections for continuous account and role hopping.

At the same time, the attacker created backdoors and mechanisms for durable persistence, such as adding new IAM users and access keys, deploying reverse shells on compute instances and containers, and subtly modifying deployment files to ensure repeated access during routine operations.

Despite the speed, all observed behaviors aligned with known MITRE ATT&CK techniques, including:

• Credential access (T1552, T1528)
• Cloud account abuse (T1078)
• Discovery (T1087, T1580)
• Command execution (T1059)
• CI/CD pipeline compromise (T1677)

Data access and extortion preparation were central to the attacker’s objectives. They executed hundreds of tailored SQL queries across multiple RDS databases to extract user records, transaction data, and other sensitive business information that could be monetized or used as leverage.

Instead of traditional ransomware-style encryption, the campaign focused on demonstrating control over infrastructure:

• Disabling S3 access
• Scaling ECS services to zero
• Blocking network access via ACLs
• Purging message queues

While these actions were mostly reversible, they indicated the attacker’s capacity to disrupt operations on demand, bolstering their position in extortion negotiations.

Several artifacts indicated that large language models directly supported the intrusion. Investigators observed highly parallel activity, with multiple AWS access keys from different accounts being used simultaneously from a single IP address and user agent.

This pattern was strongly consistent with automated, centrally orchestrated workflows rather than a single operator executing tasks manually.

Scripts and payloads contained markers of AI-generated code, suggesting that the attacker could quickly generate task-specific tools for new surfaces, such as custom discovery scripts or data extraction utilities, on demand, rather than relying solely on pre-built tools.

Attacker-created branches, commits, and HTML files often framed the activity as a “red team” or “pentest.” This language likely aimed to mislead defenders and prompt AI tools into considering the activity as authorized security testing.

What made this incident particularly notable was the imbalance between offensive and defensive uses of AI. The attacker utilized AI to maintain operational memory across numerous identities, track which credentials unlocked which resources, and coordinate overlapping waves of discovery, abuse, and persistence at machine speed.

Defenders, in contrast, struggled with fragmented visibility, partial control deployment, and manual approval processes that could not keep up with automated credential theft, permission analysis, and attack-path mapping.

This case underscores a blunt reality for cloud security teams: When AI can turn stolen credentials into a full cloud compromise within days or even minutes, resilience relies less on novel point solutions and more on robust identity security, secrets management, CI/CD hardening, and automated containment that can operate at a similar pace.

Interact with Cyber Threats in Windows, Linux, macOS VMs to Trigger Full Attack Chain - Analyse Malware & Phishing with ANY RUN