BlackGate — a Quantum Clarity research project

AI agents increasingly take actions that are fully authorized but still unsafe. Authorization tools can't see that gap. BlackGate studies how to check agent actions for safety before they run, not just after.

Authorized and safe: a two-by-two matrix A two-by-two matrix with authorized or not authorized on the horizontal axis and safe or unsafe on the vertical axis. The authorized but unsafe quadrant is highlighted in deep blue. Authorized Not authorized Safe Unsafe Authorized and safe Routine, permitted work Not authorized and safe Blocked by authorization Authorized but unsafe The gap BlackGate studies Not authorized and unsafe Blocked, and unsafe
Authorization and safety are different questions. The highlighted quadrant — authorized but unsafe — is the gap BlackGate studies.

Current work

We've pre-registered a test of whether one class of unsafe agent action can be caught before it runs, including the conditions under which we'll declare the idea dead.

Research collaboration or design-partner inquiries: [email protected]