Writing

On October 18 I launched Innerwall — an open-source microsegmentation platform for Linux VMs, bare metal and cloud instances.

Six parts treated the host's firewall as a trusted executor. Part 7, the last, of Microsegmentation from First Principles opens the box: one owned table replaced in one kernel transaction, sets that carry the churn, connection marks that never misattribute, policy on disk before it reaches the kernel, and evidence gaps that close the loop the series opened.

A first enforcement release could constrain both directions; Innerwall's constrains one. Part 6 of Microsegmentation from First Principles: why inbound is the half that pays, why outbound is a different problem rather than a smaller one, and how the other half is reserved in the contract instead of bolted on later.

Every stream is mutually authenticated, so every agent needs a certificate it got without a human and keeps without one. Part 5 of Microsegmentation from First Principles: token-gated enrollment with no trust-on-first-use window, a service boundary drawn on authentication, day-long credentials in place of revocation, and renewal as the critical path.

A rendered policy in a database protects nothing until every host holds it, exactly. Part 4 of Microsegmentation from First Principles: why policy moves as desired state, why every session starts from a snapshot, what a per-workload version has to mean, and why a distribution failure costs convergence but never safety.

A rule that names addresses is a claim about assignments, and assignments drift. Part 3 of Microsegmentation from First Principles: why policy should name workloads by label, what makes an identity safe to write policy against, and how addresses become state the control plane derives.

A candidate segmentation policy is a hypothesis about every flow an application needs, and observation cannot see the flows that didn't run. Part 2 of Microsegmentation from First Principles: why the hypothesis has to be tested as the exact ruleset enforcement will install, and what a verdict of safe to enforce has to be made of.

A segmentation policy is an allowlist under default deny, which makes it a claim about every connection an application needs. Part 1 of Microsegmentation from First Principles: why that claim has to start from observation, and what it takes to observe correctly.

You wouldn't deploy a web app without a load balancer. Why are you calling LLM APIs directly? The case for an AI gateway as essential infrastructure.

Most teams underestimate their LLM costs by 40-60%. Here's what nobody tells you about the true cost of running LLMs in production.