The researcher deployed approximately 100 AI agents on his own infrastructure to test his personal account security. The experiment lasted five hours and utilized three abliterated models: GLM-5.3, GLM-5.3 Flash, and DeepSeek V4 Flash, all running on cloud GPUs.
The agents successfully compromised three accounts by exploiting vulnerabilities in the researcher's legacy pet projects and demo applications. These exploits involved IDOR vulnerabilities and credential management errors rather than zero-day flaws in third-party software.
Beyond account access, the agents conducted 16 social engineering attempts and gathered sensitive personal information, including the researcher's home address and phone number, by cross-referencing public data and social media profiles.
The cost of the experiment was based on five hours of usage for three models on two B300 GPUs. With five successful account compromises achieved, the cost per successful breach was approximately $40, though the author estimates this could drop below $5 within a year.
