A sophisticated cyberattack orchestrated by near-autonomous AI agents has infiltrated government networks in Asia, compromising 85 employee accounts and pilfering over 2,500 personnel records. This breach, unveiled by Dream researchers, underscores the growing capability of AI frameworks to autonomously execute complex intrusion operations at remarkable speed.
Details of the Cyberattack
Investigators from Dream uncovered a 160 MB archive containing 1,395 files related to the operation, detailing 12 separate attack waves that were conducted over a span of four days in early July 2026. Utilizing the Hermes and OpenClaw frameworks, up to eight sub-agents executed tasks such as reconnaissance, credential attacks, API analysis, data collection, and lateral movement within the targeted systems.
The attack’s origin remains unconfirmed, but linguistic analysis of the operational documents suggests involvement of a Chinese-speaking entity. Although public speculation points to Taiwan as the primary target, Dream’s report refers only to unspecified government entities in Asia.
Technical Execution and Vulnerabilities
The AI-driven operation commenced by analyzing JavaScript bundles from a government portal, extracting crucial API endpoints and authentication data. These steps allowed the agents to map 21 interconnected government systems, including a single sign-on (SSO) infrastructure, revealing 36 API endpoints related to account management and user data.
The agents exploited several unauthenticated endpoints to gather employee information, such as names and department details, while also leveraging publicly available SSO integration documentation. Further analysis identified multiple server-side vulnerabilities, including insecure authentication endpoints and weak token validation, as primary contributors to the data breach.
Implications and Security Measures
The campaign demonstrated multiple attack vectors, including an automated password spray against an office automation portal. The agents employed optical character recognition (OCR) to bypass CAPTCHA security and utilized predictable password patterns to successfully compromise 85 accounts. Notably, 84 of these accounts gained access to internal systems via an SSO bridge.
Despite attempts to upload a web shell via an unrestricted file-upload endpoint, a secondary security layer prevented its execution. Nevertheless, attackers succeeded in extracting 2,564 personnel records, including sensitive employee and user information, along with internal network data and database credentials.
This incident highlights the increasing sophistication of AI agents in conducting coordinated cyber operations. To mitigate similar threats, organizations must prioritize securing exposed APIs, reinforcing SSO trust relationships, validating JWTs, and implementing comprehensive credential-spraying defenses.
The evolving threat landscape necessitates rapid investigation and response measures. By integrating threat intelligence from extensive Security Operations Centers (SOCs), defenders can enhance their Tier 1 capabilities and preemptively address potential vulnerabilities.
