Security, fraud and financial-crime teams face high volumes of alerts, fragmented evidence and adversaries who adapt quickly. Agents can collect context, query approved systems, assemble timelines and prepare cases for review. They can also amplify mistakes, expose sensitive evidence or take disruptive action on the basis of manipulated input.
This course shows how to design supervised agents for investigative work. Participants will combine deterministic detection and case-management controls with agentic reasoning where flexible evidence gathering adds value. Every significant conclusion must remain traceable to source evidence, and consequential actions require explicit authority.
The course also examines AI-enabled threats, including more scalable social engineering, malicious content and attacks against the agents themselves.
Learning Outcomes
Upon completion of this course, participants will be able to:
- Identify investigation tasks suited to supervised agent assistance
- Design bounded tools for security, fraud and AML evidence gathering
- Preserve provenance, case integrity and analyst accountability
- Evaluate agent conclusions against verified source evidence
- Manage false positives, uncertainty and escalation
- Defend agents against prompt injection and evidence poisoning
- Apply human approval to containment and customer-impacting actions
- Establish operational and governance controls for production deployment
Course Outline
The Opportunity and Threat Landscape
- Agent assistance across cyber, fraud and financial-crime operations
- AI-enabled phishing, impersonation and attack automation
- Differences between detection, investigation and response authority
- Probabilistic reasoning within regulated case workflows
- Risks of false confidence and automated customer harm
- Selecting high-value, bounded starting use cases
Designing an Investigation Agent
- Defining the case objective and success criteria
- Read-only evidence gathering before action
- Tools for logs, transactions, identity and case systems
- Structured observations, hypotheses and evidence references
- State and checkpoints across long-running investigations
- Clear stopping and escalation conditions
Alert Triage and Enrichment
- Consolidating alerts from several detection systems
- Gathering asset, customer and transaction context
- Deduplicating related alerts and identifying common entities
- Prioritising by risk, impact and evidence strength
- Explaining prioritisation without inventing causality
- Measuring whether triage reduces workload without hiding risk
Cybersecurity Investigation Workflows
- Building timelines from logs and security telemetry
- Mapping indicators across endpoints, identity and cloud services
- Query generation within approved search boundaries
- Malware and vulnerability context from trusted sources
- Drafting incident summaries for analyst review
- Escalating containment actions through authorised playbooks
Fraud and Transaction Investigation
- Customer, device, account and payment context
- Behavioural anomalies and linked-entity analysis
- Distinguishing unusual activity from evidence of fraud
- Handling real-time decisions and delayed investigation
- Customer friction, fairness and false-positive impact
- Capturing rationale and evidence for case decisions
AML and Financial-Crime Cases
- Supporting transaction monitoring and alert investigation
- Entity resolution and network-based evidence
- Typologies, scenarios and changing criminal behaviour
- Drafting case narratives from verified observations
- Separating suspicion, evidence and final determination
- Protecting sensitive intelligence and reporting processes
Evidence Integrity and Grounding
- Source provenance, timestamps and chain of custody
- Distinguishing observed facts from agent inference
- Citation and replay of queries used to gather evidence
- Conflicting, incomplete and stale information
- Preventing fabricated evidence or unsupported links
- Requirements for analyst review and case sign-off
Security of the Agent Workflow
- Prompt injection through messages, documents and case notes
- Poisoned threat intelligence and malicious external content
- Least-privilege access to tools and sensitive data
- Isolation between cases, customers and jurisdictions
- Preventing exfiltration through outputs and tool calls
- Adversarial testing of the complete investigation path
Response, Approval and Accountability
- Human approval for blocking, freezing and containment
- Separation of investigation and action authority
- Parameter-bound approvals and transaction limits
- Appeals, overrides and corrective action
- Complete logs of evidence, decisions and side effects
- Maintaining accountability under assisted decision-making
Evaluation and Production Operations
- Historical case replay and representative scenario sets
- Precision, recall, analyst time and outcome measures
- Evaluating summaries, evidence selection and policy compliance
- Monitoring drift, tool failures and changing typologies
- Incident response for unsafe or compromised agents
- Model, prompt and tool change governance
Practical Capstone
- Design a supervised investigation workflow
- Connect bounded evidence tools and case state
- Run a mixed cyber or fraud scenario
- Verify provenance and challenge unsupported conclusions
- Apply human approval to a consequential action
- Present evaluation results, controls and residual risks
Intended audience
This course is designed for security engineers, fraud and AML analysts, data scientists, software developers, security-operations teams and technical risk leaders in financial institutions. It is suitable for mixed teams designing or governing AI-assisted investigation workflows.
Prerequisites
Those attending this course should meet the following:
- Foundational understanding of cybersecurity, fraud or financial-crime operations
- Basic familiarity with data analysis and case-management workflows
- Awareness of large language models and AI agents
- Python or API experience is helpful for technical labs but not required for all participants