Three Months of InnerWarden, July to October
Three months of InnerWarden, an AI agent security startup: 70 stable releases, Community 1.0.0 to 1.5.5, release attestation, and what comes next.
InnerWarden has a small, legible architecture. The free Community guard is open source (Apache-2.0, at inner-warden); the paid stack (Pro and Enterprise) is source-available: a deterministic sensor, an interpretive agent, and a CLI, each doing one job. These posts are how it is built and how to build on it.
The sensor stays deterministic with no LLM in the hot path; the agent is the only layer allowed to reason and call out. That split is deliberate, it keeps detection trustworthy and response auditable. These posts cover the design decisions, the module boundaries, and how to extend it.
Three months of InnerWarden, an AI agent security startup: 70 stable releases, Community 1.0.0 to 1.5.5, release attestation, and what comes next.
A first-person account of why we built InnerWarden: the pain that pushed us out of MDR contracts and DIY Wazuh stacks, and what we wanted a security agent on our own hosts to do.
The story of how glibc malloc fragmentation caused our Rust daemon to grow to 1.3GB under bot traffic, and how jemalloc fixed it with 3 lines of code.
Lessons from shipping 74 eBPF programs (defined in source) across tracepoints, kprobes, LSM and XDP: #![no_std], CO-RE/BTF, ring buffers, verifier failures, debugging.
5 rounds of adversary emulation with Caldera v5.3.0. From 36% detection to 67% of testable techniques. The argv truncation fix, 15+ false positive cleanups, and why mapped is not detected.
AI codegen made shipping faster. Attacker tooling made exploitation faster. The old ship-first-harden-later loop is dead. The case for security that defaults on.
A security tool that ships with auto-block on locks operators out and gets uninstalled. The case for shipping in observe mode: detect everything, log everything, escalate on confidence.
Network IDS vs cloud APIs vs eBPF on host. Encrypted traffic, runtime visibility, fleet-scale config audit. When you'd choose each, not a settled debate.