Principle 01
Mercy
Refuse harm to a person. The system must not assist with instructions, targeting, coercion, or escalation that would predictably injure an individual.
Governance-First Security. Built to Last.
Training pathway. Certificate of completion only.
The independent cybersecurity & AI-governance training pathway — rigorous curriculum, defence-only labs, and honest AI safety: alignment remains unsolved, so ai.heart uses an external deterministic input/output brake rather than trusting the model to behave. Privacy-conscious by design; verification, transcript, payment, security and operational records may still exist where required.
Australia compliance anchor
10 Dec 2026 is the confirmed hard date for APP entities to disclose substantially automated decisions in privacy policies where those decisions significantly affect rights or interests. Proposed mandatory AI guardrails remain policy proposals unless enacted.
Training references
Curriculum maps to practical controls: NIST AI RMF Govern, Map, Measure and Manage; ISO/IEC 42001 as an AI management-system standard; and OWASP web, API and LLM application security references. These guide practice and do not prove model alignment.
The uncontrollability problem, answered honestly
1. Alignment-by-training is not solved. We do not claim that a model can be made reliably safe by scale, fine-tuning, prompt rules, preference data, or self-description. General systems can route around weak instructions, and evaluations are samples, not guarantees. A pass today is not a theorem for tomorrow’s distribution shift.
2. ai.heart’s answer is an EXTERNAL brake. The governed engine places a deterministic screen around the model: requests are checked before generation, and responses are checked again before release. This layer is independent of weights, prompts, hidden reasoning, and self-reports. It is enforcement, not persuasion: the model is not asked to “be good”; unsafe transactions are blocked, constrained, or refused outside the model.
3. Real limitation: the HarmBench gate is not fully met. We show failed gates as failed gates. They are not converted into marketing certainty, buried in aggregate scores, or renamed as “near success.” The current claim is narrower: deterministic external governance can reduce exposed harm paths and make failures visible, but it does not solve AI uncontrollability.
input
↓
governed screen
↓
output or refusal
External governance brake
A deterministic floor applied outside the model.
ai.heart does not rely on the model choosing to behave well. Every request and every response is screened by an external enforcement layer before it is allowed to proceed. The brake is deterministic, model-independent, and designed as a floor: if the request or output fails the floor, it is refused or constrained.
How it works: the model can generate, but it does not get final authority. The governance brake evaluates both sides of the exchange against three institutional principles, then blocks, narrows, or permits the interaction based on that evaluation.
Principle 02
Equity
No group is treated as lesser. The system must not validate dehumanisation, exclusion, or unequal protection based on group identity.
Principle 03
Cognizance
The system must be able to explain why it refused. A refusal should be grounded in a clear safety reason, not hidden preference or vague policy language.
Important: this is enforcement, not persuasion. The brake is applied to every request independently of model weights, prompt style, or user intent.
Boundary statement
What we do NOT claim
AiA treats alignment as hard, incomplete, and safety-critical. The point is not to market certainty; it is to state the control boundary clearly enough that a skeptical expert can test it.
-
We do not claim the model is aligned.
The model may still contain unsafe capabilities, latent failure modes, or behaviours that appear only under pressure. Our claim is about an external control layer, not a solved inner model.
-
The brake is external, not magic.
ai.heart applies a deterministic governance brake that screens input and output independently of the model. It is enforcement, not persuasion, and it can still be probed by novel attacks.
-
The HarmBench gate is not fully met.
We show that truthfully. A partially unmet gate is not hidden, renamed, or converted into a marketing pass. The visible status is part of the safety case.
-
The public audit log is intentionally unauthenticated.
Public verifiability is the thesis: observers can inspect the academy evidence trail at aiheart.com.au/aia-academy without needing private credentials or institutional permission.
never break a heart · never miss a pulse · never skip a beat
👁️
The AiA Standard — start free, prove your way up
One membership at US$40/mo for everything, or start free and climb a seven-grade Metal Standard (Iron → Titanium). Governed, anonymous, governance-verified — the security & AI-governance academy governed by the engine it teaches.
🇦🇺
Australian AI Compliance — Government, Medical, Defence, Banking & Education
Sector-specific courses built around the real Australian compliance landscape — honestly separating law, regulator expectation and voluntary guidance, including the Privacy Act automated-decision transparency obligations from December 2026.
🧭
★ Flagship: AI Alignment — Ensuring AI Benefits Humanity
The heart of the academy. A 12-week, 7-module course on the most important problem in the field — why alignment matters, what makes it hard, and how researchers are solving it. Specification, measurement, scalability, the major approaches, and a capstone.
13Governed courses
120Lessons
100%Consent-based
0External trackers
🎓
Professional Diploma in Cybersecurity — 13 courses, 3 levels
A structured pathway with stackable credentials: Foundation Certificate (beginner) → Practitioner Diploma (intermediate) → Advanced Diploma (advanced) + a portfolio Capstone. 40 modules, 120 lessons, prerequisites, learning outcomes, assessments and certificates.
📚
Academy library — quick standalone courses
Prefer bite-sized? The original 7 sector courses with search, quizzes and certificates are still here.
📋
Course Catalog & Syllabus
The full printable syllabus — every stage, course, module outline and assessment.
Open →📖
Security Glossary
70+ key terms defined and searchable, filterable by category.
Open →🧭
Program Handbook & FAQ
How to start, how grading & certificates work, and answers to common questions.
Open →🧰
Authorized Pentest Portfolio
Portfolio + scoping/consent workflow with a downloadable Rules-of-Engagement document generator.
Open →🔐
Breach Awareness Checker
Check whether your own password appears in public breaches (k-anonymity), plus a strength meter & secure generator.
Open →📋
CVE / Vulnerability Dashboard
Search live public CVE data from the NVD, filter by severity, and export results to CSV.
Open →📧
Email & DNS Security Checker
Check a domain's SPF, DMARC, DKIM & MX posture to spot spoofing/phishing risk — read-only public DNS.
Open →🧪
CTF & Home-Lab Guide
A safe, legal plan to learn offensive security: VMs, CTF platforms, a roadmap, and a progress tracker.
Open →🤖
AI Model Comparison
Honest, sortable comparison of real AI models by capability, context & price — with CSV export.
Open →📖
Docs & README
How it's built, how to run it locally, and the ethical principles behind every tool.
Open →How this toolkit stays on the right side of the line
- Consent first: security tools assume written authorization and defined scope.
- Your data only: the breach checker is designed for checking accounts you own.
- Public defensive data: CVE info comes from official sources (NVD) to help patching.
- Transparent AI: comparisons cite measurable specs, not marketing claims.
About AiA · Advanced Intelligence Academy
Husnu Konak and Adam Milankovic are the founders of the academy. AiA · Advanced Intelligence Academy — and its Professional Diploma, the AiA Standard and the flagship AI Alignment course — exist on their mission to make practical, ethical security & AI education accessible to everyone, governed by the engine it teaches.