01
Shinobi · log
Stealth monitoring. Scans inputs, logs everything, blocks nothing. Safe warm-up mode.
The bouncer. A bi-directional LLM firewall with four posture modes — observe, block inbound, sanitise outbound, or both. Deployed as a sidecar, a proxy, or an SDK middleware.
六SIX PILLARS
Each pillar is a detector, a mode, or a surface the module exposes. All six are shipped in alpha; all six are documented, tested, and extendable.
Stealth monitoring. Scans inputs, logs everything, blocks nothing. Safe warm-up mode.
Active defence. Scans and blocks suspicious inputs before they reach the LLM.
Aggressive defence. Scans and blocks dangerous LLM outputs — redact secrets, block exfil, scrub PHI.
Full protection. Inputs AND outputs with blocking on both sides. Default for regulated workloads.
Per-route rules, severity gates. (Preview — declarative policy language on roadmap; today modes configure via DEFAULT_GUARD_CONFIG.)
Sidecar, HTTP proxy, or SDK middleware — same policy, any path.
令GUARD MODES
Hattori Guard is a first-class CLI and SDK. Wire it into CI, run it ad-hoc, or shell it into a sidecar. Every invocation is reproducible and logged.
$ hattori mode samurai --routes /chat▸ inbound block on /chat active▸ policy v12 loaded▸ blocked 14 / 1,204 (1.2% at 90% coverage)
活LIVE STATE
Sample run shown under Samurai mode. Panels empty gracefully until you deploy a guard.
Log only. Warm-up.
Block inbound.
Sanitise outbound.
Block both doors.
六MODE COVERAGE
盾SAMPLE RUN
列BLOCK FEED
律PER-RULE GAUGES
役DRIFT BLOCKED
服HATTORI · GALLERY
Captures from Hattori — guard-mode toggle, per-rule gauges, and policy editor.
道OTHER DISCIPLINES