0% found this document useful (0 votes)
5 views42 pages

SecAI+_Week2

The document outlines the CompTIA SecAI+ course, focusing on securing AI systems through various modules including basic AI concepts, AI-assisted security, and governance. It details AI threat modeling processes, risks, and security controls necessary for protecting AI systems throughout their lifecycle. Additionally, it emphasizes the importance of collaboration across teams and compliance with regulations while implementing AI security measures.

Uploaded by

Kedist Tekle
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
5 views42 pages

SecAI+_Week2

The document outlines the CompTIA SecAI+ course, focusing on securing AI systems through various modules including basic AI concepts, AI-assisted security, and governance. It details AI threat modeling processes, risks, and security controls necessary for protecting AI systems throughout their lifecycle. Additionally, it emphasizes the importance of collaboration across teams and compliance with regulations while implementing AI security measures.

Uploaded by

Kedist Tekle
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

Introduction to CompTIA SecAI+

Securing AI systems

Attention Attendees:
Remember to type your messages to all panellists and attendees
Course Structure

Week / Module 1
Basic AI concepts related to cyber security

Week / Module 2
Securing AI systems

Week / Module 3
AI-assisted security

Week / Module 4
AI governance, risk and compliance

Attention Attendees:
Remember to type your messages to all panellists and attendees
AI threat modelling
• Structured approach to identify and analyze risks in AI systems
• Extends traditional threat modeling
• Data, models, prompts, and AI pipelines
• Considers all AI lifecycle phases
• Outputs
• Prioritized threats
• Recommended security controls

Attention Attendees:
Remember to type your messages to all panellists and attendees
AI-specific risks and threats
• New attack surfaces and attack types
• Model behavior or training risks
• Prompts and conversations
• Plugins, tools, and external APIs
• Embeddings and vector stores
• Sensitive data leaks and information bias risks
• Stay ahead of evolving AI-centric threats

Attention Attendees:
Remember to type your messages to all panellists and attendees
Threat resources
• Known AI Vulnerabilities and Incidents
• Real-time attack patterns and misuse
• Research on emerging AI risks and defenses
• Map Risks Across the AI Lifecycle
• Data collection, training, deployment, user interaction
• Key Resources
• MIT AI Risk Repository
• CVE AI Workgroup
• AI Vulnerability Database (AVID)
• AI Incident Database (AIID)
• arXiv

Attention Attendees:
Remember to type your messages to all panellists and attendees
Pre-requisites for AI threat modelling

Knowledge and Policies


• AI and ML Architecture • Organizational Security Policies
• Data sources and ingestion • Data classification and retention
• Training and fine-tuning pipelines • Acceptable AI use policies
• Inference endpoints and APIs • Secure development standards
• AI-Specific Threats and Frameworks • Regional Regulations and Compliance
• OWASP Top 10 • GDPR
• OWASP Machine Learning (ML) Top 10
• MITRE ATLAS
• NIST AI Risk Management Framework

Attention Attendees:
Remember to type your messages to all panellists and attendees
Pre-requisites for AI threat modelling

Teams and Context


• Coordinate Across Teams Business Context
• Security • End user characteristics
• Application architecture • Data collected and processed
• Development and DevOps / infrastructure • Data storage and processing
• AI architects / ML engineers Traditional verse AI-Driven Process
• Legal and compliance • Adjust workflows and controls

Attention Attendees:
Remember to type your messages to all panellists and attendees
Pre-requisites for AI threat modelling

Case Study
• Sector: Retail automotive platform
• AI bot assists users to buy and sell cars
• Guides transactions
• Estimates resale value from car and engine photos
• Old Process
• Human customer service desk
• Manual assessment and assigned representative
• AI-Driven Process
• Bot handles most customer interaction
• Less direct human oversight

Attention Attendees:
Remember to type your messages to all panellists and attendees
Pre-requisites for AI threat modelling

Data focus
Training Data Data Handling
• Car types and features • Data storage
• Engine manufacturing details • Encryption in transit and at rest
• Historical resale values • Input validation
• User Inputs • Integrity checks
• Car and engine photos Compliance Considerations
• Vehicle identification numbers • Personal and financial data
• Driving licenses • Example: GDPR and AI-specific rules in the
• Lease agreements EU

Attention Attendees:
Remember to type your messages to all panellists and attendees
Process of AI threat modeling

AI Threat Modeling Process


1. Gather Prerequisites and Assets
• Business use case and process flows
• Architecture diagrams and data flows
• Data used, collected, and processed
• Who uses the system and where is it deployed
• Most critical assets
• Personal/Regulated data
• Trained Model/Decisions
2. Select AI Threat Modeling Framework(s)
• OWASP Top 10 for LLMs and ML
• MITRE ATLAS
• NIST AI RMF
• STRIDE, DREAD, MAESTRO

Attention Attendees:
Remember to type your messages to all panellists and attendees
Process of AI threat modeling

AI Threat Modeling Process


3. Identify and Categorize Threats
• Think like an attacker
• Consider prompt injection, poisoning, jailbreaks, model theft, etc.
• Map threats to chosen frameworks
4. Analyze and Prioritize Threats
• Assess likelihood of attack
• Evaluate business and legal impact
• Rank threats by severity

Attention Attendees:
Remember to type your messages to all panellists and attendees
Process of AI threat modeling

AI Threat Modeling Process


5. Recommend and Document Controls
• Create threat modeling report
• Map threats to mitigations and owners
• Collaborate with stakeholders on remediation plan
6. Validate and iterate across software development life cycle (SDLC)
• Security testing and assessments at each stage
• Re-run threat modeling after major changes
• Monitor for new AI-specific threats over time

Attention Attendees:
Remember to type your messages to all panellists and attendees
Threat Modeling Frameworks
OWASP Top 10 for LLM

Attention Attendees:
Remember to type your messages to all panellists and attendees
OWASP Top 10 for LLM

Attention Attendees:
Remember to type your messages to all panellists and attendees
OWASP Top 10 for ML

Attention Attendees:
Remember to type your messages to all panellists and attendees
OWASP Top 10 for ML

Attention Attendees:
Remember to type your messages to all panellists and attendees
Threat modeling frameworks
MITRE Adversarial Threat Landscape for AI Systems (ATLAS)
• Knowledge base of tactics/techniques targeting AI/ML
• Focuses: data poisoning, model evasion
• Maps attacks to phases of ML pipeline
• Supports creating detection/defense strategies

Attention Attendees:
Remember to type your messages to all panellists and attendees
Threat modeling frameworks
NIST AI Risk Management Framework (AI RMF)
• Guides organizations in understanding/managing AI risks
• Promotes trustworthy/responsible AI
• Safety, security, fairness, privacy, explainability, resilience
• Built around four core functions
• Map, measure, mange, govern

Attention Attendees:
Remember to type your messages to all panellists and attendees
Overview of Security Controls
Security Controls
• AI brings automation benefits and new attack surfaces
• Controls cover entire AI lifecycle
• Data collection and sanitization
• Model development and fine-tuning
• Deployment and integration
• Logging, monitoring, and auditing
• Focus on tech security, but also
• Access control and validation
• Misuse prevention and ethical use
• Compliance and governance

Attention Attendees:
Remember to type your messages to all panellists and attendees
Security Controls
Layered AI Security
• Model Controls
• Protect model behavior and assets
• Input validation, output filtering, watermarking, limited exposure
• Gateway Controls
• Middle layer between users and models
• Strict access control, usage limits, logging, observability
• Guardrail Testing and Validation
• Define what the AI can/can’t do
• Policies/rules enforced on inputs, outputs, and context
• Goal: Keep AI trustworthy, observable, and hard to abuse

Attention Attendees:
Remember to type your messages to all panellists and attendees
Security Controls
Layered AI Security: Gateway Layer

Attention Attendees:
Remember to type your messages to all panellists and attendees
Model Specific Controls
• Functions of Security Controls
• Manage model behavior and risk
• Protect training data, parameters, and outputs
• Examples
• Input validation and sanitization
• Output content filtering and moderation
• Rate limitation
• Model watermarking and fingerprinting
• Goal
• Keep model within defined risk tolerances across its lifecycle

Attention Attendees:
Remember to type your messages to all panellists and attendees
Model Specific Controls
On-premise models
• Full control over data, training and infrastructure
• Evaluate data lineage, access controls, environmental hardening

Attention Attendees:
Remember to type your messages to all panellists and attendees
Model Specific Controls
Pre-Trained Models
• Verify training data integrity and provenance
• Assess embedded bias and harmful content

Attention Attendees:
Remember to type your messages to all panellists and attendees
Model Specific Controls
Vendor Delivered Models
• Audit vendor security assurances and certifications
• Validate contractual and regulatory compliance

Attention Attendees:
Remember to type your messages to all panellists and attendees
Model Specific Controls
• Model security controls
• Maliciously crafted prompts or data
• Attempts to manipulate behavior or cause misclassification
• Tricking the model into sharing sensitive information or parameters
• Reverse-engineering
• Evaluating repeated querying
• Critical for proprietary data or models
• Harmful output evaluation
• Harmful, offensive, or non-compliant responses
• Content moderation for open-source, crowd-sourced data

Attention Attendees:
Remember to type your messages to all panellists and attendees
Model Specific Controls
Model Evaluation
Trustworthiness Testing Infrastructure Controls
• Consistent and transparent behavior • Authentication, role-based permissions,
session management
• Alignment with ethical, legal, organizational
• API security controls
expectations Baseline Security Review
• Access and Rate Controls • Guardrails
• Rate limitation to restrict request frequency • Red teaming exercises
• Prevents prompt chaining, brute-force
attacks, resource exhaustion

Attention Attendees:
Remember to type your messages to all panellists and attendees
Model Specific Controls

Attention Attendees:
Remember to type your messages to all panellists and attendees
Model Guardrails
• Guardrails are structured controls around model inputs/outputs.
• Implemented as code-based frameworks
• Validate and regulate prompts and responses
• Enforce security and organizational policies
• Enhance model reliability
• Prevention of unsafe or harmful behaviors
• Reduction of data leakage and policy violations
• More predictable and trustworthy model behavior

Attention Attendees:
Remember to type your messages to all panellists and attendees
Model Guardrails

Attention Attendees:
Remember to type your messages to all panellists and attendees
Model Guardrails

Attention Attendees:
Remember to type your messages to all panellists and attendees
Model Guardrails

Attention Attendees:
Remember to type your messages to all panellists and attendees
Model Guardrails
• Essential control in AI systems • Custom guardrails
• Ensures • Tailored to organization
• Safety • Risk profiles
• Compliance • Regulatory requirements
• Responsible usage • Business use cases

Attention Attendees:
Remember to type your messages to all panellists and attendees
Prompt Template
Predefined structures for LLM prompts
• Define
• System instructions
• User input placeholder
• Contextual data sections
Benefits
• Consistent model behavior and tone
• Controlled insertion of user data
• Separation between system and user prompts
Example
• “You are a secure assistant. Answer the following user question honestly and briefly:
{{user_input}}"

Attention Attendees:
Remember to type your messages to all panellists and attendees
Prompt Template
Security Benefits
• Protect against prompt injection
• User input stays in user context, not system instructions
• Prevent system prompt leakage
• System instructions never echoed to users
• Support context isolation
• User cannot change core policy or configuration via prompt
• Increase guardrail effectiveness
• Clear boundaries for where to apply input and output checks

Attention Attendees:
Remember to type your messages to all panellists and attendees
Gateway and Interface Controls
• AI gateway: Specialized API gateway for AI traffic
• Sits between user interfaces and AI models
• Key functions
• Enforce usage rules and security policies
• Route, inspect, and log AI request and responses
• Provide observability and metrics on AI usage and abuse
• Adds AI-specific safeguards
• Prompt filtering and moderation
• Token management and limits
• Modality restrictions (text, image, audio, files)

Attention Attendees:
Remember to type your messages to all panellists and attendees
Gateway and Interface Controls

Attention Attendees:
Remember to type your messages to all panellists and attendees
Questions

Attention Attendees:
Remember to type your messages to all panellists and attendees
Coming up next week.

Week / Module 1
Basic AI concepts related to cyber security

Week / Module 2
Securing AI systems

Week / Module 3
AI-assisted security

Week / Module 4
AI governance, risk and compliance

Attention Attendees:
Remember to type your messages to all panellists and attendees
References

• CompTIA (2026). CertMaster Perform SecAI+.

Attention Attendees:
Remember to type your messages to all panellists and attendees

You might also like