State-Sponsored Hackers Deploy Multiple AI Agents in Campaign Targeting Taiwanese Government Systems
Between July 1 and 4, 2026, a sophisticated campaign targeting Taiwanese government entities leveraged up to eight AI-driven sub-agents with minimal human interaction. The attack framework, based on Hermes and OpenClaw agents, executed 12 waves of intrusion attempts, cracking 85 employee credentials and exfiltrating over 2,500 personnel records while deploying persistent backdoors. Initial reconnaissance exploited JavaScript decompilation from a government portal to map 21 interconnected systems and the national single sign-on (SSO) infrastructure, revealing weaknesses such as JWT tokens accepted with the algorithm set to "none" and vulnerabilities exploitable via password spray attacks enhanced by OCR technology. The autonomous agents conducted Bayesian vulnerability scoring, iterative learning cycles, and self-managed false positive reduction, while circumventing AI model refusals by posing as authorized penetration tests. The attribution remains uncertain, although Chinese-language operational notes and regional focus suggest likely involvement of a state-backed actor from Asia. Recorded Future classifies this campaign as demonstrating Level 4 AI Malware Maturity, indicating multi-step planning with limited human oversight. Analysts recommend enhancing API authentication, removing debug endpoints, implementing strict access controls, continuous monitoring for exposed AI subscriptions, and auditing third-party AI components to mitigate similar threats.
Linha do tempo
-
State-Sponsored Hackers Deploy Multiple AI Agents in Campaign Targeting Taiwanese Government Systems
ler a análise desta fonte ocultar
Between July 1 and 4, 2026, a sophisticated campaign targeting Taiwanese government entities leveraged up to eight AI-driven sub-agents with minimal human interaction. The attack framework, based on Hermes and OpenClaw agents, executed 12 waves of intrusion attempts, cracking 85 employee credentials and exfiltrating over 2,500 personnel records while deploying persistent backdoors. Initial reconnaissance exploited JavaScript decompilation from a government portal to map 21 interconnected systems and the national single sign-on (SSO) infrastructure, revealing weaknesses such as JWT tokens accepted with the algorithm set to "none" and vulnerabilities exploitable via password spray attacks enhanced by OCR technology. The autonomous agents conducted Bayesian vulnerability scoring, iterative learning cycles, and self-managed false positive reduction, while circumventing AI model refusals by posing as authorized penetration tests. The attribution remains uncertain, although Chinese-language operational notes and regional focus suggest likely involvement of a state-backed actor from Asia. Recorded Future classifies this campaign as demonstrating Level 4 AI Malware Maturity, indicating multi-step planning with limited human oversight. Analysts recommend enhancing API authentication, removing debug endpoints, implementing strict access controls, continuous monitoring for exposed AI subscriptions, and auditing third-party AI components to mitigate similar threats.