PUBLIC SUMMARY – This is an abbreviated overview of the CAIC Standard™. The full Implementation Specification is available under a commercial license.
Controlled AI Core™ (CAIC) is a normative security standard for autonomous AI systems. It provides an architectural control layer that enforces autonomy limits, human accountability, and immutable audit—turning AI safety from policy promises into measurable, enforceable, and legally‑binding requirements.
The CAIC Standard™ defines a practical, auditable, and human‑accountable architecture for deploying AI systems with controlled autonomy, explicit responsibility, and enforceable safety mechanisms.
As AI systems transition from assistive tools to autonomous and agentic systems, organizations face growing risks related to:
- Loss of human control
- Undefined responsibility and accountability
- Regulatory non‑compliance
- Unsafe or irreversible autonomous actions
The CAIC Standard™ addresses these risks by providing a vendor‑neutral governance architecture that makes AI systems controllable, auditable, and legally defensible.
The CAIC Standard™ is:
- A governance and control standard, not a software product
- Model‑agnostic and vendor‑independent
- Designed for enterprise and high‑risk AI systems
- Focused on operational control, not ethics statements or principles
- Built around human responsibility by design
It is not: a software product, a regulatory framework, or a certification guarantee.
The CAIC Standard™ is structured in three layers:
- Governance Layer – autonomy scale (L0–L4), risk classification, lifecycle management, and human accountability (AISO).
- Technical Control Layer – runtime enforcement via kernel‑level constraints, including a reference eBPF implementation for Linux.
- Operational Security Layer – continuous behaviour monitoring, adaptive defence, supply chain integrity, and mandatory coverage of OWASP LLM Top 10 and MITRE ATLAS.
The standard is built around several key mechanisms:
- Control Plane – a mandatory interception layer between users and AI models.
- Autonomy Levels (L0–L4) – explicit limits on AI decision‑making authority.
- Policy Engine – machine‑enforceable rules defining allowed and forbidden actions.
- Pre‑ and Post‑Inference Gates – control points evaluating requests and responses before and after model execution.
- AISO (AI System Owner) – a formally accountable human role responsible for AI system behaviour.
- Self‑Attestation – regular health checks of security components performed by the customer.
- Lethal Trifecta – automatic lowering of autonomy to L2 when untrusted inputs, sensitive data access, and external tools are simultaneously present.
- Dynamic Autonomy Ceiling – automatic step‑down of the autonomy level upon detecting deviations (drift, monitoring failures, etc.).
- Kill‑Switch & Incident Control – mandatory emergency shutdown and incident response mechanisms.
- Immutable Audit Logging – verifiable records of all AI decisions and control actions.
The full version of the standard includes 26 normative protocols, structured as:
- 16 Governance protocols (A1–A15, A5a) – autonomy, policies, testing, audit, roles, lifecycle.
- 6 Technical Control protocols (A16–A19, A23–A25) – runtime security, MCP gateway, agent control, RAG security, identity & access management, secret & key management, API security.
- 4 Operational Security protocols (A20–A22, A26) – observability, continuous monitoring, supply chain integrity, resilience & fault tolerance.
- Standard Version: v2.5 (current)
- Publication Status: Public summary available; full specification under commercial license
- Certification: Available for licensees upon request
- Domain‑specific policies: Not included in this public summary
Controlled AI Core™ and CAIC Standard™ are original works.
- Use of this public summary is governed by the terms defined in LICENSE.md.
- Trademark usage is governed by TRADEMARK.md.
- Commercial use, certification, and redistribution of the full specification require explicit written permission from the author.
Albert Shamin LinkedIn
Independent author and maintainer
The CAIC Standard™ is provided as a governance and architectural standard.
It does not constitute legal advice, regulatory approval, or a warranty of compliance.
Organizations remain fully responsible for the lawful and safe operation of their AI systems.
For licensing, certification, or access to the full Implementation Specification, please contact the author directly.
© 2026 Albert Shamin. All rights reserved.
CAIC Standard™, Controlled AI Core™ – trademarks pending.