AI Agents Orchestrate Sophisticated Government Breach, Stealing 2,500+ Records
A coordinated cyberattack leveraging eight AI agents successfully breached Asian government systems, compromised 85 accounts, and exfiltrated over 2,500 personnel records.
A sophisticated cyberattack, orchestrated by up to eight autonomous AI agents, has successfully breached government systems in Asia, compromising 85 accounts and stealing more than 2,500 personnel records. The campaign, documented by researchers at Dream, utilized open-source AI agent frameworks like Hermes and OpenClaw to automate significant portions of the intrusion lifecycle, operating at machine speed.
The AI-driven framework initiated its operation by meticulously analyzing JavaScript bundles from a government portal. This reconnaissance phase allowed the agents to map 21 connected government systems, including single sign-on (SSO) infrastructure, by extracting API endpoints, OAuth client IDs, and authentication details. One target system exposed over 36 API endpoints for account management and user data, some of which were unauthenticated, providing direct access to employee information such as names, departments, and SSO account identifiers.
The confirmed compromises were primarily attributed to server-side vulnerabilities, including unauthenticated APIs, insecure authentication endpoints, and weak token validation. The agents exploited hidden API endpoints that returned valid authenticated sessions without requiring credentials. Additionally, automated password spraying attacks were conducted against an office automation portal, with CAPTCHA images being solved using OCR. This multi-pronged approach led to the cracking of 85 accounts across several attack waves.
A particularly concerning vulnerability involved JSON Web Token (JWT) processing, where an API accepted tokens with the 'none' algorithm. This allowed attackers to forge authentication tokens without possessing a valid signing key. The compromised accounts were then tested against connected internal services, achieving a remarkable 98.8% success rate in authenticating to an internal information system via an SSO bridge, granting access to internal dashboards and administrative tools.
While not all automated attack steps were successful, with a web shell upload attempt being blocked by a secondary authentication layer, the attackers managed to exfiltrate a significant volume of sensitive data. This included at least 2,564 personnel records, internal network ranges, SSO client secrets, and database credentials. The operation, which occurred between July 1 and July 4, 2026, involved 12 distinct attack waves.
What distinguishes this campaign is its advanced use of feedback loops and machine learning. The AI agents generated structured reports after each wave, ranked attack paths using Bayesian probability scoring, and initiated 'Learning Cycles' to search public vulnerability databases when initial techniques failed. The framework also demonstrated the ability to detect and discard false positives, indicating a high degree of operational sophistication.
Researchers noted that internal reporting within the campaign's documentation used Simplified Chinese, while target analysis employed Traditional Chinese, suggesting a Chinese-language operator. Public reporting has identified Taiwan as a potential target, though Dream's report broadly refers to victims as government entities in Asia.
This incident underscores the growing threat posed by AI agents in executing complex cyber intrusions. Defenders are urged to prioritize securing exposed APIs, strengthening SSO trust relationships, rigorously validating JWTs, implementing robust credential-spraying controls, and monitoring public-facing debug endpoints to mitigate similar attacks.