Reflections
Book a Demo
Mirror Reflections · Research and experiments
Mirror protects both sides of an AI system. Encryption keeps prompts, context, memory and evaluations private across supported paths, including FHE inference. Goal-bound authority limits what agents can delegate, access and commit. Each investigation shows one of these boundaries with primary sources and runnable tests.
05 investigations · Updated 1 Sep 2026
A valid opaque reasoning artifact can be replayed without breaking its encryption. Mirror replaces the portable artifact with a handle bound to the original user, session, and model.
Zero Data Retention limits residual copies after processing. It does not stop conventional inference infrastructure from receiving usable plaintext while the model runs.
One encrypted model pass is not a response. Token selection, KV memory, feedback, ciphertext schema, precision, refresh, GPU placement, and output release must remain connected across the complete supported generation.
AVERI and Google DeepMind showed that private benchmarks and proprietary model weights can meet inside an attested enclave. This note shows how FHE can reduce the plaintext and trusted-code surface inside that shared boundary.
A capable agent can recover from a blocked path by finding another route. The security question is whether that new information can create new authority. We classify six incident patterns and test the same task with and without destination-verified action control.
Encrypt intelligence before it crosses a boundary. Keep agent authority tied to the original goal. Verify every consequence before it becomes real.