Understand this story
In an interview, Ben Horowitz recounted an agent security testing story: the team had one agent build a website for a pet store, and the resulting code was riddled with vulnerabilities; they then had another agent launch an attack, and the latter immediately exploited the vulnerabilities to break in and steal all the customer data. He noted that the vendor's response to the incident was "you wrote your prompts wrong," blaming an engineering flaw on user skill. Ben Horowitz believes that once an agent has the permissions to call tools, read and write databases, and execute autonomously, relying on prompts to prevent risk is like running around naked, and the only solution is to weld permission isolation and circuit breakers into the architecture at the lowest level. The related report was published on October 18, 2026.
AI-generated from public reports
Latest development · Report
Ben Horowitz on Agent Security and the Vendor Responsibility VacuumDevelopments
1 confirmed developments
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 18
AYi AI Notes
Ben Horowitz on Agent Security and the Vendor Responsibility Vacuum
In an interview, Ben Horowitz recounted: when his team tested top closed-source models, they had one agent build a website for a pet store, and the code was riddled with flaws; then they had another agent launch an attack, and the latter immediately exploited the vulnerabilities to break in and take all the customer data. The vendor's response to the incident was "you wrote your prompts wrong," blaming an engineering defect on user skill. He believes that once an agent has the authority to call tools, read and write databases, and execute autonomously, relying on prompts to prevent risk is like running around naked, and the only solution is to weld permission isolation and circuit breakers into the architecture at the lowest level.
ayi_ai_notes
English translation is being prepared.