Deadlight

Offline AI safety and execution for the Navy's AI-first fleet.

As frontier AI moves from the cloud toward ships and autonomous systems, connectivity cannot be assumed. SATCOM can be degraded or intentionally denied through EMCON, contested spectrum, physical obstruction, weather, or network outages. When the uplink disappears, what keeps an AI agent running and safe offline?

Deadlight is a offline sandboxed harness that sits between frontier models — OpenAI, Grok, Claude, or local models — and the systems they can affect. Agents propose actions and Deadlight decides whether they can execute.

Technical Architecture

• Daytona Sandboxes — isolated high fidelity execution simulations with filesystem, process, tool, and network controls.
• Vector Embeddings — high-dimensional local representations of mission context, telemetry, policies, and historical states for offline semantic retrieval.
• Signed Policy Snapshots — versioned local policies with capability-level permissions, TTLs, and fail-closed behavior.
• Candidate-State Evaluation — evaluates the predicted result of an agent's action before committing it using the policies and the high fidelity sandboxed environment.
• Safety Gates — combines deterministic policy enforcement with local anomaly detection to catch both known violations and unexpected behavior.
• Model-Agnostic Orchestration — the underlying model can change without changing the safety boundary.

Our prototype demonstrates an agent operating normally with an uplink, then continuing safely after SATCOM is cut. Unsafe actions—such as violating an EMCON policy or leaving an authorized corridor—are blocked before execution, leaving the system at its last known-safe state.

Deadlight is not a replacement for Navy C2 or autonomy systems. It is the trust and execution layer between increasingly capable AI and the physical world.

The AI-first fleet cannot be cloud-dependent.

Offline. Sandboxed. Policy-bound.