By Mirror Security · SOC 2 Type 1 Certified · Free Access

AI Security Training for
Teams Building with AI

A structured knowledge base covering every layer of the AI attack surface. Four tracks from threat fundamentals to hands-on labs with VectaX, AgentIQ, and DiscoveR.

Risk Assessment →
4 tracks
39 modules
OWASP · NIST AI RMF · MITRE ATLAS
EU AI Act · ISO 42001
Jump to track:
Track 1: Fundamentals
Track 2: Attack Surface
Track 3: Defence in Depth
Track 4: Applied Labs

Mirror Security

Platform & products

Documentation-style pages for each product: encrypted AI with VectaX, runtime guardrails with AgentIQ, and automated red teaming with DiscoveR.

FHE & vectorsVectaXEncrypted vector memory, fully homomorphic encryption Search, and encrypted inference for RAG and ML pipelines.View VectaX →Agents & policiesAgentIQRuntime safety: prompt and PII checks, policy engine, tool-call controls, and compliance-oriented guardrails.View AgentIQ →Red teamingDiscoveRAutomated adversarial testing for LLMs and agents — injections, jailbreaks, RAG attacks, and CI-friendly scans.View DiscoveR →

How it works

Four tracks. One complete picture.

Each track builds on the last. Every topic appears exactly once. Click any module to see what's inside and open it directly.

1
Orientation

AI Security Fundamentals

Threat landscape, frameworks, and threat modeling

5/5 modules available

AI Security Fundamentals

5 modules

Five modules from first principles to OWASP Top 10 and threat modeling.

01What AI security is and why it differs from traditional security02The AI threat landscape: attack categories, real incidents, MITRE ATLAS03OWASP Top 10 for LLMs: all 10 risks with examples04AI threat modeling: applying STRIDE and MITRE ATLAS to a real system05AI governance and compliance: NIST AI RMF, EU AI Act, ISO 42001, GDPR
2
Attack Surface

Three domains of AI attack surface

Pick the path relevant to your stack, or do all three

Parallel paths18/18 modules available

A: RAG and Vector DB Security

6 modules

How RAG pipelines are built and how they fail.

A1Architecture: how RAG and vector DBs workA2RAG attack surface: poisoning, inversion, injection via retrievalA3Securing the embedding pipelineA4Access control, identity, and data governanceA5Encrypted inference and vector memoryA6RAG security in production
VectaXDiscoveRreferenced where relevant

B: AI Agent Security

6 modules

How AI agents fail and how to secure agentic systems.

B1Agent architecture and how agents failB2Prompt injection: direct, indirect, multi-stepB3Tool use and MCP attack surfaceB4Input/output guardrails and bypass techniquesB5Agent identity, least privilege, credential scopingB6Multi-agent trust and orchestration risk
AgentIQreferenced where relevant

C: Model and Training Attacks

6 modules

Model evasion, extraction, poisoning, and MLOps security.

C1Model evasion: adversarial examples and classifier bypassC2Model extraction and membership inferenceC3Training data poisoning and backdoor modelsC4Supply chain: model registries, deps, AI-BOMC5Secure MLOps: CI/CD pipeline integrity for AIC6AI red team methodology: plan, execute, report
DiscoveRreferenced in C6
3
Defence in Depth

Cross-cutting concerns

Privacy, governance, and security operations

Parallel paths10/10 modules available

D: Privacy-Preserving AI

5 modules

Cryptographic and mathematical tools for protecting data through the AI lifecycle.

D1Why AI privacy differs: inference attacks, reconstructionD2FHE deep dive: PHE, SHE, FHE, CKKS, noise, bootstrappingD3Differential privacy: mechanisms and applying DP to AID4Federated learning: architecture, threat model, poisoningD5SMPC: secure multiparty computation for AI
VectaXD2 references the cryptography behind VectaX

E: Security Operations for AI

5 modules

Detection, monitoring, incident response, and compliance for AI deployments.

E1Zero trust architecture for AI systemsE2Security monitoring and anomaly detection across AI stacksE3AI incident response: playbooks, forensics, containmentE4Compliance in practice: NIST AI RMF, ISO 42001, EU AI ActE5AI as an attack tool: adversarial AI and AI-powered threats
4
Applied

Industry scenarios with Mirror Security

The only track where Mirror products are the subject

6/6 modules available

F: Healthcare and Life Sciences

2 modules

Encrypted AI on patient data. HIPAA compliance from the stack up.

F1Encrypted AI inference on PHI with VectaX: full buildF2HIPAA compliance validation: audit trail and key management
VectaX

G: Financial Services

2 modules

Secure RAG for trading and fraud detection. End-to-end red teaming with DiscoveR.

G1Secure RAG for trading and fraud detection with VectaXG2Red team a financial AI system end-to-end with DiscoveR
VectaXDiscoveR

H: Enterprise Agentic AI

2 modules

AgentIQ policy design for production agents. Sovereign AI deployment.

H1Design and enforce an AgentIQ policy for a production agentH2Sovereign AI deployment: FIPS 140-3, air-gapped FHE
AgentIQVectaX

Who it's for

Built for the people closest to the problem

Three roles, different entry points, same knowledge base.

Security Engineers

Find and fix AI-specific risks

Attack surfaces unique to LLMs, agents, and RAG pipelines. Test for prompt injection, data leakage, and model manipulation before they become production incidents.

Track 1 → 2A, 2B, 2C → 3E
AI / ML Engineers

Build defensibly from the start

Security principles that fit the architecture before deployment. Secure model selection, inference endpoint hardening, agent tool permissioning, encrypted memory.

Track 1 → 2A, 2B → 3D → Track 4
Compliance and Risk Teams

Map AI systems to your frameworks

Practical guidance on GDPR, EU AI Act, ISO 42001, and NIST AI RMF with controls that actually apply to LLMs, agents, and generative AI deployments.

Track 1 → 3E (E4 especially)
Aligned with
OWASP Top 10 for LLMsMITRE ATLASNIST AI RMFEU AI ActISO / IEC 42001GDPR & HIPAA

FAQ

What practitioners ask most

Do I need to complete the tracks in order?
Track 1 is the recommended starting point for everyone. Within Track 2, the three paths (RAG, Agents, Model attacks) are independent. Track 3 and 4 assume Track 2 knowledge but not a specific sequence.
Why does the same topic not appear in multiple paths?
Every topic has exactly one home. Prompt injection lives in path B, FHE fundamentals live in path D, compliance lives in Track 1 and 3E. No topic appears twice.
What is the OWASP Top 10 for LLMs?
The OWASP Top 10 for LLM Applications (2025) catalogues the ten most critical security risks in generative AI: Prompt Injection, Sensitive Information Disclosure, Supply Chain, Data Poisoning, Improper Output Handling, Excessive Agency, System Prompt Leakage, Vector Weaknesses, Misinformation, and Unbounded Consumption. Track 1 module 3 covers all ten.
Where do Mirror Security products appear?
VectaX, AgentIQ, and DiscoveR appear in Tracks 1 to 3 only where genuinely relevant. Track 4 is the only track where they are the subject of the learning.
What is MITRE ATLAS?
MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) is a knowledge base of adversary tactics and techniques targeting AI and ML systems. As of v5.1 (November 2025) it contains 16 tactics and 84 techniques. Track 1 module 2 introduces it; module 4 teaches you to apply it.
Is this free?
Yes. Mirror Academy is a free public knowledge base. No registration required. Track 4 applied labs reference Mirror Security product playgrounds which have free tiers at docs.mirrorsecurity.io/mirrordocs.

The threats keep moving.
Get ahead of them.

AI security cannot be an afterthought. Start with the fundamentals and build up to hands-on red teaming.

Get a Risk Assessment