Securing the Path to Superintelligence
Understanding the vulnerabilities in our current systems is the first step toward building a secure future.
This report is intended to provide a new threat model and a roadmap for securing superintelligence. As we continue to build more powerful AI systems, we must develop security that protects us from both external actors wielding AI and rogue, cyber-capable models attacking from within. Solving this requires defending against sabotage, escape, and theft (SET):
S: We must ensure that the models trusted to red-team and patch our software have not themselves been sabotaged.
E: We must prevent models from breaking containment, executing commands on external systems, or exfiltrating their own weights. We must also be able to locate and shut down untethered models.
T: We must protect model weights from sophisticated threat actors seeking to steal the most value-dense digital assets ever created.
If we can build the right security, run advanced AI safely, and trust the systems we deploy, the benefits will be immense. Superintelligence could compress centuries of scientific and medical progress into years, curing diseases, uncovering new mathematics, and solving problems humanity has never been able to touch.
Now that we have seen the warning signs and can identify the emerging vulnerabilities, we must take up the sacred mission of building the protections needed to secure these systems before their capabilities outrun our defenses.
We must secure the path to superintelligence. Let’s begin.