Four pillars. Fourteen dimensions. One operating model for AI governance — mapped to APRA CPS 230, CPS 234, AS ISO/IEC 42001, ASD ISM and AICD director-duty guidance.
Australian boards are now formally accountable for AI under AICD director-duty guidance, APRA CPS 230, and AS ISO/IEC 42001. Existing AI frameworks define what to do — but rarely in a form a board can table, an auditor can defend, or an operator can use.
The Mycelium 14-Dimension AI Security Framework was authored to fill that gap. One operating model. Three practices. Evidence at every dimension.
The same four pillars sit behind every Mycelium Health Check, every Agent Build, and every Board Briefing.
Each dimension carries a control intent, an evidence expectation, and a clear pillar home. Together they form the scoring surface used in the AI Security Health Check.
The structural pillar. Names who is accountable, what the appetite is, and how AI is governed across its lifecycle.
AI strategy aligned to business strategy, with board-endorsed objectives, scope and explicit risk appetite.
Named board-level accountability for AI risk, with formal charter, RACI and reporting cadence.
Explicit AI risk appetite, tolerance thresholds and escalation triggers, integrated with enterprise risk.
Lifecycle governance for AI models — approval gates, change control, performance review and decommissioning.
AI risk assessed at vendor onboarding and on annual re-certification cycle. Embedded AI explicitly covered.
The security pillar. Protects the data, the models, the access pathways and the upstream supply chain that feeds AI systems.
Data lineage, classification, retention and protection for AI training and inference data.
Protection against model tampering, data poisoning, model extraction and prompt injection.
Role-based access, least-privilege defaults, secrets management and segmentation for AI systems.
Provenance and integrity of models, libraries, datasets and tooling across the AI supply chain.
The operational pillar. Covers the live-running discipline around AI systems — what you see, what you do when something breaks, how you change it safely.
Drift detection, performance baselines, anomaly alerting and audit logging on production AI systems.
AI-specific incident playbooks, kill-switch authority and tested rollback procedures.
Controlled release process for AI model updates — documented testing, approval gates and rollback paths.
The human pillar. Without literacy and a healthy reporting culture, every other control degrades over time.
Role-specific AI literacy, including board-level training refreshed on regulatory change.
Organisation-wide AI awareness, responsible use protocols and a healthy reporting culture.
Every dimension carries an explicit mapping to the Australian and international standards already in force or imminent. The framework is the bridge from your AI posture to your regulatory posture.
In force. Names AI as material technology requiring board oversight and risk management. Anchors accountability, model governance and incident response.
In force. Establishes board accountability for information security capabilities, including AI systems. Anchors the Protect pillar end-to-end.
Now an Australian standard. Defines management-system requirements for organisations using AI. Maps across all four pillars — the framework operationalises it for Australian boards.
Five major publications since 2024. Directly ties director duties of care and diligence to AI literacy, accountability and governance.
Expands obligations around automated decision-making, model transparency and consumer rights. Anchors data governance and accountability obligations.
The international reference. The Mycelium framework is structurally compatible with NIST AI RMF (Govern/Map/Measure/Manage) but mapped to Australian regulatory expectations.
Australian Government Information Security Manual update. New AI-specific controls (ISM-2112 through 2123) covering human approval gates for risky actions, behavioural baselines and drift detection, shadow AI prevention, and secure prompt/output deletion. IRAP assessors will start demanding evidence against named controls.
For Internal Audit, IRAP assessors and board sub-committees who need the trace in one walk. Each map shows where every named clause sits in the 14 dimensions — PRIMARY where the control lives natively, SUPPORTS where the dimension co-delivers it.
The new AI controls (ISM-2112 through 2123) mapped to the four pillars. For government, defence, critical infrastructure and IRAP-bound buyers.
Download as PDF ↓
CPS 220, 230 §§24-33, CPS 230 §§40-60, CPS 234, CPS 510 + CPG mapped. For CROs and CISOs at banks, insurers and superfunds defending controls to APRA.
Download as PDF ↓
GOVERN, MAP, MEASURE and MANAGE functions mapped to the 14 dimensions. For boards comparing global frameworks and cross-border organisations.
Download as PDF ↓
All ten LLM-specific threats — prompt injection through unbounded consumption — mapped to the dimensions that govern them. For CISOs and AppSec teams.
Download as PDF ↓Want to talk one through? Book a diagnostic call ↗
The framework is the operating model behind every Mycelium engagement. Each practice uses it differently — but the dimensions never move.
A 4-week diagnostic that RAG-scores your organisation across all 14 dimensions. The framework provides the scoring surface, the evidence checklist and the standards mapping.
See the Health Check ↗Every agent is assessed at build time against the framework. The Approval Gate, Audit Log and Owner roles in the architecture map directly to D04, D10 and D02.
See the agent practice ↗Board literacy and posture briefings are anchored to the framework. Directors learn the four pillars; the board reads its posture against the 14 dimensions.
Download the Boardroom Diagnostic ↗Every dimension is scored RAG against evidence, not opinion. The same scoring discipline is used across Health Check, Agent Build, and Board Briefing.
Control is documented, owned, operating, and the evidence holds up to auditor review. Refreshed on the documented cadence.
Control is partially in place, or in place but not validated against the standard. Evidence exists but is incomplete, stale, or not consolidated.
Control is absent or fundamentally inadequate. Likely material exposure under CPS 230, ISO 42001 or director-duty obligations. Priority remediation item.
Each dimension expects evidence of one or more types: documented policy, operating artefact, system record, training record, tested procedure, or third-party attestation. The Detailed Findings Report names the evidence reviewed and the gaps observed, dimension by dimension.
The framework is a living document. Updates are versioned, dated, and communicated to active clients. The Mycelium team maintains it against regulatory change so your posture stays current.
Initial public version. Four pillars established (Govern / Protect / Operate / Enable). Fourteen dimensions defined. Standards mapping established for APRA CPS 230, CPS 234, AS ISO/IEC 42001, ASD ISM (AI controls 2112-2123), IRAP, AICD director-duty guidance, Privacy Act 1988 (ADM 2026).
RAG scoring methodology codified. Evidence types defined. Refresh cadence published.
Next review scheduled for September 2026 in line with anticipated APRA AI prudential standard release.
Thirty minutes. No pitch. No proposal until it makes sense.
Book a diagnostic