Three primitives.
Built into the kernel.
The Hook. The Sentry. The Gate. One proprietary engine, attached below the agent, in the host kernel, via Aya-compiled eBPF. Read the spec.
The Nvidia Hook
Total visibility into the black box of GPU compute. Microsecond intercept on every CUDA launch and every ioctl.
The Hook attaches below the agent — in the host kernel — via Aya-compiled eBPF. A uprobe on libcuda.so:cudaLaunchKernel records every GPU launch. The syscalls:sys_enter_ioctl tracepoint records every driver-bound ioctl. The agent doesn't see us. The driver doesn't see us. We see everything.
The Zombie Sentry
Stop paying for AI that isn't thinking. Save up to 20% on GPU bills by killing hung agents in real time.
Forgotten notebooks. Runaway loops. Agents stuck firing the same kernel a thousand times a second while the bill clock runs. We watch the kernel-launch fingerprint of every PID through the eBPF ring buffer. When a process crosses the threshold, we alert — and if policy says so, the kernel signals SIGKILL within milliseconds. Your CFO gets the GPU-hours back.
The Exfiltration Gate
Phi-3 powered intelligence analyzing intent at the driver level. Your PII never leaves the server.
Every suspect ioctl gets two passes. Stage one is a kernel-side heuristic on the hot path: size estimate plus reservoir sample, sub-microsecond. Stage two hands the survivors to an on-host Phi-3 mini that scores intent against a learned exfil profile and returns a 0–100 risk rating with a reason. The model runs on the same host as the workload — no cloud round-trip, no third-party API. Your weights, your customer PII, your competitive data: never out the door.