Wiz Research analyzes real-world cases and trajectories of AI agent attacks
Start here
In incident response and threat hunting, Wiz Research observed that in multiple real intrusions, AI agents appeared to carry out most of the hands-on operations and may have orchestrated the attacks themselves. Public cases include an ARTEX agent penetration-testing framework reported by researchers to have attacked a South Korean financial institution and exfiltrated the personal information of approximately 67,000 people, the ConnectWise supply chain incident that Microsoft attributed to Storm-0940 (attempted logins to more than 2,250 tenants), and activity of the Hermes agent discovered by Hunt.io in the intrusion at Thailand's Ministry of Finance. Wiz Research also outlines methods for identifying the characteristics of such attacks and discusses their implications for the future of cybersecurity.
Generated by AI from reports · Updated 14 days ago
Related research
Papers whose work the reports mention
- The reports mention PentestGPTPentestGPT: An LLM-empowered Automatic Penetration Testing ToolThe rise of agentic attacks: what they look like and where they are heading
arXiv · 2023-08-13
Progress timeline
- WebBloomberg Technology· Media coverageThe Cyber Race Goes Agentic(opens the original post in a new tab)
- Wiz ResearchWiz Research: The rise of agentic attacks—real cases, identifying characteristics and future direction
AI agents are changing how attackers operate; Wiz Research reviews real cases of agentic attacks that have emerged and gives methods for identifying the characteristics of such attacks. The article also discusses the impact of such attacks on the future of cybersecurity.
- Wiz ResearchThe rise of agentic attacks: what they look like and where they are heading
Wiz Research observed in incident response and threat hunting that in multiple real intrusions, AI agents appeared to carry out most of the hands-on operations and may have orchestrated the attacks themselves. Public cases include an ARTEX agent penetration-testing framework attack reported by researchers against a South Korean financial institution, exfiltrating the personal information of approximately 67,000 people; a ConnectWise supply chain incident that Microsoft attributed to Storm-0940 (attempted logins to more than 2,250 tenants); and Hunt.io's discovery of the Hermes agent operating in unattended mode in an intrusion at Thailand's Ministry of Finance. The three attacks Wiz investigated entered via an exposed Flowise instance, a Java gateway RCE, and a GraphQL backend with an MCP interface, respectively; one of them generated more than 150 scripts in about 21 hours, targeting about 25 cryptocurrency exchanges and fintech companies, and stole cloud credentials and Kubernetes tokens.
Related discussion
- Older than 72 hours · Not included in the 72-hour figures
- WebBloomberg TechnologyThe Cyber Race Goes Agentic(opens the original post in a new tab)