AI Risk Explorer (AIRE)

A lone Taiwan-based operator ran 12 AI agents to find targets, write exploit scripts and exfiltrate data from organizations in 11 countries. At least 10 were breached, exposing patient-care data from Taipei’s fire department and 119 emergency dispatch records. Source: Acronis. How it compares with the rest of our Threat Watch: - It's the 3rd most sophisticated of 26 cyber operations run by an individual, and in the top 6% of all 475 cyber operations. - It's the first operation in our database where abliterated open-weight models (DeepSeek, GLM and Qwen) ran complete intrusions unattended. Commercial models, including Grok and China’s MiniMax, built much of the infrastructure but refused to attack. - Nine AI tools named, more than double any other operation in our database, mixing coding assistants, agent frameworks and offensive security tools. - Seven sectors hit, the broadest targeting by any individual in our database. Of the eight cyber operations in our database that hit six or more sectors, seven ran AI agents through multi-step intrusions: agentic campaigns make breadth cheap.

We’ve updated our Threat Watch, which tracks real-world cyber operations and influence campaigns where AI contributed to the operation. Looking at the cyber data, we see signs of a shift on five fronts: capabilities, models, actors, victims, and reporting. A thread 🧵

August in AI risk, from our latest newsletter: