Audit first. Enforce when you’re ready.

The Guardrail eBPF beta is a two-tier rollout designed around one idea: you should never be surprised by your own security layer.

Tier 1 — See everything

Deploy in audit mode. Nothing is blocked. Your real workload runs unchanged, and every syscall that policy would block is logged with blocked:1 — the prediction. You get a complete picture of what enforcement will do on your machines, with zero risk of breaking a workflow.

Tier 2 — Block, deterministically

After a clean audit record and an explicit sign-off, flip the same policy to enforce. The denials you saw predicted are the denials you get — audit predictions and enforce denials are computed by the same policy logic, verified 1:1 in our dry runs. There is no model guessing what to block: the policy you wrote is the only thing that decides. If a command is allowed in audit, it stays allowed in enforce.

What Guardrail enforces

• Exec of deny-listed commands (exact basename, comm-level)
• Mounts, ptrace, kernel module loads, and competing BPF program loads — per-UID policy
• UID allowlists: root and trusted accounts bypass cleanly
• Seven LSM hooks + six audit tracepoints, verified live on modern kernels

Every check happens in the kernel in microseconds. No LLM in the enforcement path, no telemetry, no cloud dependency.

Eligibility

We’re looking for partners running:

• Production Linux (x86_64) with kernel-level control (BPF LSM available)
• Real workloads with a tolerance for a short audit window first — including container hosts, which get specific guidance
• A willingness to keep an audit trail (loader version + policy hash) and report what you find

Apply

Tell us about your hosts, kernel versions, and what you want to lock down:

Elai-intelligence@pm.me

Beta partners get direct access to the operator runbook, recovery procedures proven in drills, and a direct line to the engineer who built it.