GIAC AI Platform Security

Adapting a foundation model changes its behavior and its exposure. You examine fine-tuning, alignment methods and moderation controls, and the security implications of training data, backdoors and weakened safety behavior.

• Fine-tuning and parameter-efficient adaptation
• Alignment methods and why they can be undone
• Data poisoning and backdoors in training data
• Moderation layers on input and output
• Evaluating a customized model’s safety
• Controls for who may customize a model

You build A security assessment of a fine-tuning process, with a list of required controls.

Retrieval gives a model facts and gives an attacker a new route in. You review retrieval-augmented architectures, secure vector databases and ingestion pipelines, and handle access control when documents carry different permissions.

• How retrieval-augmented generation is built
• Vector databases: access, isolation and exposure
• Ingestion pipelines and untrusted documents
• Access control that respects document permissions
• Embedding leakage and inference risks
• Testing a retrieval system for injected content

You build A security review of a retrieval system, with access control tests and ingestion safeguards.

The model is one part of a larger application. You evaluate architecture patterns and development frameworks for GenAI, find supply chain weaknesses in models and libraries, and review the integration layer where tools and plugins connect.

• Reference architectures for GenAI applications
• Development frameworks and their default risks
• Model, dataset and package supply chain
• Plugins, tools and function calling as attack paths
• Secrets, logging and data handling in application code
• Architecture review checklist for GenAI apps

You build An architecture review of a GenAI application, with a risk-rated list of findings.

Agents act on their own, so their mistakes have consequences. You design and assess agentic systems, secure communication between agents, manage context safely, and review the protocols that connect agents to tools.

• Agent architectures and levels of autonomy
• Tool access and least privilege for agents
• Agent-to-agent communication and trust
• Context management and memory poisoning
• Protocol-level security for tool connection standards
• Human approval and audit trails for agent actions

You build A threat assessment of an agent system, with permission design and approval points.

Where a model runs decides who can reach it. You assess the risks of hosting models and applications in cloud and elsewhere, and choose controls for endpoints, secrets, networks and the specialized hardware underneath.

• Hosting options: managed services, self-hosted and edge
• Securing model endpoints and inference APIs
• Network isolation and private connectivity
• Secrets and credentials for AI workloads
• Rate limiting, abuse prevention and cost controls
• Hardware and accelerator security considerations

You build A deployment security baseline for a model-serving environment.

A model is only as trustworthy as the pipeline that produced it. You secure data workflows, training pipelines and the model life cycle with MLOps and MLSecOps practices that make changes reviewable and reversible.

• Data pipelines: provenance, validation and access
• Training environments and their isolation
• Model registries, versioning and signing
• CI/CD for models with security gates
• Monitoring models in production for drift and abuse
• Retiring models and their data

You build A secured ML pipeline design with provenance checks, signing and release gates.

Structured threat modeling applies to AI as it does elsewhere, with new threats added. You adapt familiar methods, draw on frameworks such as MITRE ATLAS and the OWASP lists for LLM applications, and produce a model that a team can maintain.

• Adapting STRIDE and data flow diagrams to AI
• MITRE ATLAS tactics and techniques
• OWASP guidance for LLM applications
• Threats at training, deployment and inference
• Ranking and tracking AI threats
• Keeping a threat model current

You build A threat model for a GenAI application, with each threat mapped to a mitigation.

AI use grows faster than most governance. You apply risk management frameworks such as the NIST AI RMF to AI systems, plan security over the long term, and support responsible use without making the business wait.

• AI risk management frameworks in outline
• Governance structures for AI use
• Responsible AI and transparency obligations
• Balancing security with business speed
• Roadmap and maturity for AI security
• Reporting AI risk to leadership

You build An AI security roadmap with governance, controls and reporting for one organization.

Audit one complete platform from model to pipeline. You map the system, review each layer, test the most exposed paths and write findings and a remediation plan a platform team can act on.

• Scoping and mapping the platform
• Reviewing model, retrieval and agent layers
• Reviewing infrastructure and pipelines
• Testing exposed paths
• Writing findings and a remediation plan

You build A full security audit report of a GenAI platform with prioritized remediation.