Built to what independent authorities are actually asking for

Gartner's AI Trust, Risk & Security Management framework describes governance as five layers: AI Governance (policy authored), Runtime Inspection & Enforcement, Information Governance, Infrastructure & Stack, and Traditional Protection underneath. Most tools in the category sit in one layer. Luminous's enforcement engine spans three — Runtime Enforcement, Information Governance, and Infrastructure — unifying, in one deterministic engine, controls the framework treats as separate layers: it enforces the action, governs the data the action touches, and extends into the confidential-compute substrate underneath. It integrates upward with whatever policy-authoring tools govern the first layer, and inherits standard protections from the fifth rather than replacing them.

Gartner's own research is why this gap matters: through 2026, Gartner estimates at least 80% of unauthorized AI transactions will trace back to internal policy violations, not external attacks — an authority problem, not a threat-detection problem. Gartner also requires genuine TRiSM tooling to be independent of any single model or cloud provider; Luminous is model- and cloud-agnostic by design.

OWASP's GenAI Security Project publishes a Top 10 risk list specific to agentic AI applications (ASI01–ASI10). Mapped against that list, Luminous's architecture closes 7 of the 10 risks by construction — meaning the risk is structurally unrepresentable for any action that passes through the enforcement engine, not merely detected or logged after the fact — addresses 2 more through a combination of controls, and shares responsibility for the remaining 1 (agentic supply-chain risk) with the customer's own infrastructure and dependency management, which is outside any single vendor's control.

The pattern worth naming: OWASP's own recommended mitigations for agentic risk are, almost without exception, the same authority controls Luminous is built around — an external policy engine separating planning from execution, per-action authorization instead of a standing grant, short-lived task-scoped credentials, least privilege, human approval for high-impact actions, and immutable signed records. That's not a coincidence Luminous is claiming — it's OWASP's own prescription, independently arrived at.

Joint guidance co-issued by the Five Eyes cyber authorities — CISA and NSA in the US, ASD's ACSC in Australia, the Canadian Centre for Cyber Security, NCSC-NZ, and NCSC-UK — reaches the same conclusion from a different direction. The guidance is explicit that detection-based controls are structurally weak against agentic threats: impersonated or spoofed agents, in the guidance's words, "evade audit controls" and defeat detection tuned to "identify normal behaviour," leaving anomaly detection "ineffective until a confirmed anomaly surfaces." Their prescribed fix is the same shape as OWASP's: "continuous runtime authentication with centralised policy decision points for each action," entitlements re-checked at every invocation rather than once at startup, and a specific recommendation to "quarantine requests to delete logs or audit records until reviewed and approved by a human."

Two independent authorities — a government cyber-security coalition and an application-security standards body — converged on the same control set from different starting points. That convergence is stronger evidence than either body's endorsement of any one vendor, because neither is endorsing Luminous specifically.

Independent evaluation of agentic AI systems under realistic conditions — not vendor benchmarks — has found the large majority of agents fail at the level of judgment, not capability: they know how to take an action, but nothing stops them from taking the wrong one, or taking a right action at the wrong time, with no authorization step in between. That's a governance gap, not an intelligence gap, and it's the specific gap this platform is built to close.