Install
Install in sixty seconds.
The community edition ships everything: runtime governance, KYP scoring, attack chains, DIDs, evidence chain, framework adapters. Apache 2.0, self-hosted from day one.
The AI control plane for runtime governance. KYA is a framework-agnostic governance layer for AI agents performing autonomous actions, built on five architectural primitives introduced in the KYA paper. Apache 2.0. Deployable in your environment.
The community edition ships everything: runtime governance, KYP scoring, attack chains, DIDs, evidence chain, framework adapters. Apache 2.0, self-hosted from day one.
Each primitive is independently useful. Together they form the governance layer — making every autonomous action authorized, policy-conforming, and post-hoc verifiable.
Every action passes four checks — identity, authority, policy conformance, evidence integrity — before dispatch. Sub-millisecond scorer at p99. 1,800 ops/sec at 20 workers.
Three-channel multi-tenant hierarchy: regulator → tenant → agent. Tenants can only tighten default weights, never loosen. Formal proof — no accidental permissiveness.
Schema-level unification of trust scoring across human users, AI agents, and service accounts. One table, one score, delegation-aware.
Auditable multiplier over an AIVSS-shaped additive baseline. Interactions between principals amplify or damp risk with a formal audit trail.
Static premium for risky delegates + runtime debit for actual delegate misbehavior in multi-agent fan-out. Catches topology-guided attacks.
Beyond the five architectural primitives above, the runtime kernel ships with the operator-facing controls regulated enterprises need in production from the first install — no assembly required.
Cap the cost of a bad decision before it lands. Per-agent, per-tenant, per-tool.
Halt an agent, a fleet, or a whole tenant in one click. Fires in sub-millisecond.
32 frameworks — EU AI Act, HIPAA, GDPR, SOC 2, PCI, DORA, SR 11-7, and more — mapped to articles, retention windows, and breach fan-out.
W3C Decentralized Identifiers and Verifiable Credentials for every principal — humans, agents, service accounts.
Correlates events across agents, delegations, and tenants. Catches multi-step attacks single-shot detectors miss.
Every verdict tied to an immutable kernel record — HMAC-chained, Ed25519-signed, replayable months or years later.
KYA doesn't compete with observability — it composes with it. Where LangSmith, Phoenix, Arize, and Fiddler tell you what happened, KYA tells you whether it was allowed. Bridges shipping in-repo for Fiddler, Phoenix, and any OpenInference-emitting stack.
How long the agent ran. How much it cost. What path it took. What tools it called. Great at describing.
Was this action authorized? Did it conform to policy? Can we prove it to an auditor? Great at deciding.
Every number below is reproducible against the Apache 2.0 codebase. No wall-garden benchmarks.
1,200 probes from PyRIT and Garak, including the Liang 2025 topology-guided multi-agent attack. False-positive rate reported per family.
Pure-function scorer with no I/O in the hot path. Runs in the customer process, not a network round-trip.
HMAC chain integrity preserved end-to-end under concurrent load.
PostgreSQL, MySQL, SQLite, DuckDB × 9 test scenarios. All 36 cells green. Backend-agnostic by construction.
