Secure Acceleration report details risks to AI infrastructure

A stealth AI security lab released the report *Secure Acceleration* at secureacceleration.com. It distinguishes model-behavior security, including prompt injection and alignment, from AI-infrastructure security. The report’s authors say infrastructure expansion is moving faster than security, while important models—described as approaching AGI or ASI—are being deployed on systems with limited security.

The risks outlined include model sabotage that degrades performance, sleeper agents, backdoors, data poisoning, models escaping controlled environments, and theft of model weights. A sophisticated intelligence service stealing weights is presented as a hypothetical scenario; the weights, the report’s speaker said, embody billions of dollars of investment in a few terabytes. The report also discusses cyber swarms and multi-agent systems, and says data centers were not designed to contain superintelligence.