Observed arrival · 2026-10-27
Silent Zero OS puts the agent below the safety loop
A working paper proposes a minimal, formally verified real-time OS architecture for running untrusted AI agents, with authority constrained by human-granted capabilities.
- For
- OS, verification, firmware, and robotics researchers
- Worth noticing
- The paper labels each guarantee as proven, tested, or assumed, making its assurance boundaries explicit.
Field notes
The design puts authority in explicit, scope- and time-limited capabilities, with a deterministic loop retaining final say over physical motion. It treats processor time as controlled authority too, aiming to prevent an agent from starving the control path. The authors identify composition across verified and unverified components as the central challenge; firmware and GPUs remain trusted, while verified multicore support and timing bounds are acknowledged gaps.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue