Tamper-evident audit logs, sandboxed execution, and human approval gates are table-stakes hygiene any serious agent runtime ships. What is harder to fake: an Apache 2.0 kernel anyone can read, a monthly P&L anyone can audit, and a founder whose GitHub handle is the same one that signs every commit. See /forge/benchmark for the moat audit.
Public mapping of kernel controls to the LLM Top 10.
In flight · Q3 2026
Where we ARE differentiated
Two layers competitors cannot replicate in under twelve months.
The controls on this page are necessary, not sufficient. Our defensible moat lives elsewhere: the multi-year intel graph (Polymarket priors, whale-wallet patterns, news embeddings, calibration curves) and the 6-hook deterministic inference router. Benchmarks — including the cases where we lose — are published openly with reproducible methodology.
Standard enterprise hygiene primitives — we publish ours for transparency, not as a unique differentiator. Five primitives are live in the Kairon Forge kernel right now; every claim below references a specific file in the open-source kernel or a named admin surface, so nothing is aspirational on this list.
Tamper-evident audit log
Every Guardian-mediated action — approval, agent step, MCP tool call, replay — emits an immutable, hash-chained audit row. Operators can export the chain as JSONL or CSV for SIEM ingestion; the chain itself is verifiable independently of Kairon-controlled infrastructure.
Untrusted skill-pack code executes inside a WebAssembly sandbox with bounded CPU time, bounded memory, and no ambient filesystem or network. Per-customer concurrency caps keep one tenant from starving another at the runtime layer.
Every /admin and /api/admin surface is guarded at three independent layers: the edge proxy checks admin_role on the JWT, the server-side layout calls getAdminCaller() and redirects on miss, and each per-route handler runs requireAdmin() or requireModerator(). One missing layer is caught by the other two in CI tests.
Issued API tokens are stored only as sha256 hex hashes in the kairon_os_api_tokens table. The plaintext is returned to the user exactly once at issue time. A database compromise yields hashes, not credentials; rotation is one row update.
kairon_os_api_tokens (sha256 hex hash column)
Per-RPC OpenTelemetry tracing
Every gRPC and REST call through the kernel emits a per-RPC OpenTelemetry span with tenant + caller + outcome attributes. Spans chain into the streaming + sandbox layers so a single audit row can be correlated to its underlying compute trace.
Six items on the dated procurement roadmap. Each carries a target quarter rather than a “soon” badge so a procurement reviewer can plan around the schedule and a security buyer can hold us to it.
SOC2 Type I — scope frozen
Initial scope: kairon-guardian kernel + Forge control plane + customer audit-log export path. Readiness assessment in progress; Type I report targeted for Q3 2026.
Q3 2026
Vanta Trust Center — live portal
Continuous-monitoring portal exposing policy index, sub-processor list, and control evidence to enterprise prospects without an NDA round-trip. Currently being configured.
Q3 2026
Public threat model (STRIDE)
A versioned PDF threat model covering admin gate, audit log, WASM sandbox, JWT WebSocket bridge, and the MCP tool-registration path. Public, signed, and re-issued on material kernel changes.
Q3 2026
Independent penetration test
A scoped external engagement with Trail of Bits or NCC Group against the kernel + Trust Center surface. Findings + remediation summary published.
Q4 2026
HackerOne bug bounty
Public bug bounty program with a structured severity matrix and disclosure SLA. Scope: kairon-guardian, the Forge control plane, and the customer-facing audit-log export path.
Q4 2026
SOC2 Type II — audit window
Twelve-month observation window opens once Type I is issued. Final report timing depends on the issued Type I date; current target window opens Q1 2027.
Q1 2027
Guardian Phase 3 — kernel evolution
The Trust Center roadmap above tracks procurement artifacts (SOC2, pen test, bug bounty). This section tracks the kernel itself. Four primitives ship in Phase 3 alongside the compliance work.
Persistent audit DB (Rust-backed)
Q3 2026
Replace the current in-memory + SQLite audit pool with a Rust-native persistent store (sled / rocksdb candidate) wired into the kernel event log. Tamper-evident hash chains are preserved across the boundary; the read-model projection into Supabase guardian_logs remains the customer-visible surface.
services/kairon-guardian/src/bridge/audit_db.rs
gRPC streaming dual-transport
Q3 2026
First-class streaming on the gRPC + REST dual transport: server-streaming for audit-tail subscriptions, client-streaming for batched event ingest, and bidi for the sandbox execution channel. Both transports remain bound to :9000 with the same Bearer auth and per-IP rate-limit middleware.
Runtime verification that an installed skill pack carries the intel-graph scopes its manifest declares. On install + on each agent step the kernel checks the user tier against required_intel_scopes; missing scopes return 402 with the upgrade hint and never silently degrade behaviour.
Audit-log root anchors and skill-pack publisher signatures move to hardware-key signing (YubiHSM 2 candidate). Each audit checkpoint and each marketplace pack release ships with a verifiable signature that does not depend on a Kairon-controlled software key.
EU Representative engagement is in flight. Until a named EU Representative is contracted, GDPR Art 27 / EU AI Act inquiries route through eurep@kairon.trade. Acknowledged within two business days. Target contracting quarter: Q3 2026.
Replay-from-checkpoint, native sandbox, RBAC/MFA, rate limiting, and OpenAPI generation are commodity B2B hygiene any team can ship. Read our honest moat audit for what makes Kairon Forge genuinely different.
Kairon AI
Kairon AI
Powered by local AI. Upgrade to Pro for Claude API access.