Amazon Threat Intelligence Uncovers Widespread Compromise of FortiGate Devices
Amazon Threat Intelligence has identified that over 600 FortiGate devices have been compromised across more than 55 countries by a financially motivated, Russian-speaking threat actor with moderate technical skills. This actor leveraged at least two distinct commercial large language model (LLM) services to automate attack planning, reconnaissance tool generation, and post-exploitation operations—capabilities that would typically require a significantly larger and more skilled team.
Notably, there were no exploits of FortiGate vulnerabilities observed during this operation. Initial access was acquired solely through exposed management interfaces—specifically ports 443, 8443, 10443, and 4443—and by exploiting weak or reused credentials with single-factor authentication. What stands out in this campaign is not the technical sophistication of the threat actor, but rather the operational scale made possible through AI. When faced with hardened environments or more advanced defensive measures, the actor would disengage from the target, opting instead to focus on less secure systems, underscoring that their advantage derived from AI-enhanced efficiency rather than advanced exploit capabilities.
Amazon Threat Intelligence reports that the actor utilized at least two different commercial LLM services for complementary operational roles. The first served as the primary tool developer, attack planner, and operational assistant, while the second acted as a movement planner when assistance was needed to navigate within specific victim networks. In one documented instance, the actor submitted the complete internal topology of an active victim to the second LLM—including IP addresses, hostnames, confirmed credentials, and identified services—requesting a detailed plan to compromise additional inaccessible systems.
The AI-generated attack methodologies featured comprehensive plans that included command sequences, estimated success rates, execution timeframes, and prioritized task trees. These plans referenced academic literature on offensive AI agents, indicating that the actor was actively monitoring publications related to AI-assisted penetration testing. However, a documented structural limitation was evident: while the actor produced technically sound sequences, they consistently failed to adapt when conditions deviated from the established plan, struggling with custom exploit compilation, debugging failed attempts, or creative movement against non-standard defenses.
Access Vector for Active Directory Directories
The actor’s staging infrastructure, left publicly accessible—a significant operational security oversight—contained AI-generated attack plans, victim configurations, and the source code of custom tools. This inadvertent exposure provided Amazon Threat Intelligence with comprehensive visibility into the methodologies used and how AI was integrated into every phase of operations.
FortiGate configuration files represent high-value targets in this attack chain as they include recoverable SSL-VPN credentials, administrative credentials, complete network topology with routing information, firewall policies revealing internal architecture, and IPsec VPN peer configurations. Once these configurations were exfiltrated, the actor exploited the stolen credentials to connect to the victims’ internal networks via VPN and carry out post-exploitation activities.
The post-exploitation methodology followed a standardized sequence: deploying Meterpreter with the mimikatz module for DCSync attacks against domain controllers, extracting NTLM hashes from Active Directory, and then executing lateral movements using hash replay and Kerberos ticket attacks. In confirmed compromises, the attacker obtained complete databases of domain credentials. In at least one documented case, the domain Administrator account utilized a plaintext password, undermined by either reuse or weak credentials extracted from the FortiGate configuration.
The actor specifically targeted Veeam Backup & Replication servers through various credential extraction tools—including PowerShell scripts, compiled decryption utilities, and attempts to exploit known vulnerabilities (CVE-2023-27532, CVE-2024-40711). The focus on backup infrastructure aligns with a systematic ransomware deployment strategy, aiming to compromise recovery capabilities before any encryption takes place. However, the actor’s operational notes indicate repeated failures against patched or uniquely configured targets.
An AI-Generated Toolkit
Analysis of the custom reconnaissance tool’s source code—available in both Go and Python—reveals hallmark characteristics of AI generation: redundant comments that merely restate function names, an overly simplistic architecture with disproportionate investment in formatting at the expense of functionality, naive JSON syntax parsing through string matching rather than proper deserialization, and empty documentation stubs. The tool automates post-VPN reconnaissance workflows: ingesting target networks from VPN routing tables, classifying by size, discovering services via the gogo port scanner, automatically identifying SMB hosts and domain controllers, and performing vulnerability assessments via Nuclei on discovered HTTP services.
The volume and variety of the custom tools—configuration analyzers, credential extraction utilities, automated VPN logins, large-scale scanning orchestration, and aggregation dashboards—typically suggest a well-resourced development team. CJ Moses, Amazon Integrated Security’s CISO, refers to the model as an “AI assembly line for cybercrime,” empowering less skilled actors to operate at scale. This entire toolkit appears to have been generated by an individual actor or a very small group using AI-assisted development without sufficient expertise to mitigate failures under unexpected conditions.
Geographical targeting seems opportunistic rather than sector-specific, consistent with automated large-scale scanning: significant concentrations of compromised equipment have been observed in South Asia, Latin America, the Caribbean, West Africa, Northern Europe, and Southeast Asia. Clusters suggest compromises at the organizational level—contiguous blocks of IP addresses or shared non-standard management ports—indicating deployments by managed service providers or large enterprise networks. For cybersecurity professionals, the geographic coverage of this campaign highlights that no region or sector is off-limits for a low-level actor operating with AI enhancements.



