Security, Compliance & Risk Management at Organizational Level

50 minutes • Video + Lab

The CIA Triad and Enterprise AI Security

Information security in enterprise environments is typically framed using the CIA triad: Confidentiality (only authorized people access information), Integrity (data hasn't been altered or corrupted), and Availability (systems and data are accessible when needed). For AI systems managing sensitive nonprofit data, all three dimensions matter.

Confidentiality breaches occur when unauthorized people access sensitive donor or beneficiary data. Integrity failures happen when data is corrupted or AI models are adversarially manipulated to produce false results. Availability failures occur when systems fail, preventing staff from accessing AI tools they depend on. Effective enterprise AI security addresses all three.

Key Takeaway

Enterprise AI security is not a technical function isolated to IT—it's an organizational responsibility spanning leadership, technology, operations, and culture. Security failures at scale affect entire organizations and erode stakeholder trust.

Threat Modeling for AI Systems

Understanding threats is the first step toward defense. AI systems face distinctive threat categories beyond traditional software.

Data Poisoning Attacks

AI models learn from training data. If an attacker can corrupt training data, the resulting model produces biased or incorrect predictions. Example: if an attacker adds fake beneficiary records with exaggerated needs to program evaluation datasets, AI models might learn to inflate severity estimates. Detection: validate training data quality, monitor for statistical anomalies in training datasets, include data validation in model training workflows.

Adversarial Examples

Attackers craft specific inputs designed to fool AI models. A charitable donor's record might be slightly modified (name variant, different address version) to manipulate propensity scores. Detection: monitor for unusual input patterns, test model robustness against perturbations, implement input validation rules.

Model Extraction Attacks

An attacker might query an AI system repeatedly to reverse-engineer the underlying model, then use that stolen model externally or for attacks. Mitigation: rate-limit API queries, monitor for unusual access patterns, contractually prohibit competitors from using systems.

Privacy Attacks (Membership Inference)

An attacker might determine whether specific individuals are in training datasets, violating privacy. If someone learns their data was used in an AI model without consent, trust erodes. Mitigation: implement differential privacy (mathematical techniques adding noise to protect individual records), limit model transparency about training data, contractually restrict external model access.

Supply Chain Attacks

Compromises of AI vendors (cloud platforms, model libraries) can cascade into your organization. Mitigation: vendor security assessment, contractual security requirements, continuous monitoring of vendor security posture.

Organizational Risk Assessment

Comprehensive risk assessment evaluates likelihood and impact of potential harms, then prioritizes mitigations.

Risk Identification

Systematic risk identification involves: threat brainstorming (what could go wrong?), vulnerability assessment (what weaknesses exist?), asset identification (what are we protecting?), and scenario development (if threats exploited vulnerabilities, what impact?). Risks span technical (database breach), operational (staff misconfiguration), process (inadequate access controls), and governance (lack of oversight).

Risk Prioritization

Not all risks deserve equal resources. Risk prioritization matrices typically plot likelihood vs. impact: high-impact, high-likelihood risks get immediate attention. High-impact, low-likelihood risks require contingency planning. Low-impact risks might be accepted. This framework ensures resources address highest risks first.

Risk Response Strategies

Mitigation: Take actions reducing likelihood or impact (encryption reduces breach impact). Acceptance: Understand risk and proceed anyway (small organizations might accept certain risks they lack resources to mitigate). Avoidance: Don't pursue initiatives with unacceptable risks. Transfer: Buy insurance to transfer financial consequences of certain risks.

Apply This

Conduct a risk assessment for a major AI initiative your organization is considering: Identify 5-10 potential threats. For each, estimate likelihood (low/medium/high) and impact (low/medium/high). Plot on 3x3 matrix. For high-priority risks (high likelihood or high impact), develop mitigation strategies. This exercise crystallizes abstract security thinking into concrete planning.

Security Controls Framework

Security controls are specific actions reducing risk. The NIST Cybersecurity Framework organizes controls across five functions.

Identify

Understand what you're protecting: asset inventory (what systems, what data), data classification (what's sensitive?), and risk assessment (what could go wrong?). You can't protect what you haven't identified.

Protect

Implement safeguards: access controls (who can access what?), encryption (protecting data in transit and at rest), secure configurations (systems hardened against common attacks), training (staff understand security responsibilities).

Detect

Identify breaches and unauthorized activity: monitoring systems for intrusions, detecting unusual access patterns, alerting to anomalies. Detection assumes breaches will occur; the question is how quickly you find them.

Respond

When incidents occur, respond effectively: incident response procedures (defined steps minimizing damage), communication protocols (who tells whom), investigation (understanding what happened), remediation (fixing the problem).

Recover

Restore normal operations: restore systems from backups, validate data integrity, implement improvements preventing recurrence.

Access Management

Access controls are foundational to security. Determine who can access what, under what conditions, and enforce consistently.

Authentication and Authorization

Authentication confirms identity (you are who you claim). Authorization determines what authenticated users can do (you can access donor records but not finance data). Multi-factor authentication (MFA)—requiring something you know (password) plus something you have (authenticator app) or something you are (biometric)—significantly improves security. Authorization should follow principle of least privilege: users get minimum access needed for their role.

Credential Management

Passwords are weak security controls—people reuse them, share them, and forget them. Better approaches: Single Sign-On (SSO) centralizing authentication across systems, eliminating password proliferation; API keys for system-to-system authentication; OAuth for delegated access. Enforce strong password policies when passwords are necessary: minimum length (12+ characters), complexity (mix of types), regular changes.

Segregation of Duties

Prevent single individuals from controlling critical operations. No one person should both approve grants and record grant payments; no one person should both modify data and audit changes. Segregation prevents fraud and catches errors.

Encryption and Data Protection

Encryption converts readable data into unreadable form using keys. Only holders of correct keys can decrypt.

Encryption in Transit

Data moving between systems (user to server, server to server) should be encrypted. HTTPS for web traffic, encrypted VPNs for remote access, encrypted database connections. This prevents eavesdropping on network traffic.

Encryption at Rest

Data stored on servers should be encrypted. If an attacker gains physical access to servers or steals disk drives, encrypted data remains unreadable without keys. Encryption keys should be managed separately from encrypted data (don't store keys on same server as encrypted data).

Key Management

Encryption security depends entirely on key security. Keys must be: generated securely (using cryptographic random number generators), stored securely (restricted access, encrypted when possible), rotated periodically (changing keys to limit damage from key compromise), and destroyed securely (when no longer needed). Key management services (AWS KMS, Azure Key Vault, Google Cloud KMS) handle this complexity for you.

Vendor Security Assessment

Most enterprise nonprofits rely on external vendors (cloud platforms, CRM vendors, AI providers). Vendor security directly affects your security.

SOC 2 Type II Certification

SOC 2 is an audit standard assessing vendor controls around security, availability, processing integrity, confidentiality, and privacy. Type I audits vendor systems at a point in time. Type II audits over a period (typically 6+ months), assessing whether controls operate consistently. Type II is more meaningful. Ask vendors for SOC 2 Type II reports. If a vendor providing financial or sensitive data management lacks SOC 2, this is a red flag.

ISO 27001 Certification

ISO 27001 is an information security management standard. Certification indicates the vendor has implemented and maintains a comprehensive information security program. This is particularly important for vendors handling sensitive data.

Vendor Due Diligence Questionnaire

Create a vendor security questionnaire asking about: security controls (encryption, access management, monitoring), compliance (SOC 2, ISO 27001, industry-specific requirements), incident response (how do they handle breaches?), data protection (how long do they retain data? where is it stored?), and business continuity (what's their backup and disaster recovery capability?). Require vendor completion and review responses seriously—they reveal vendor maturity.

Incident Response Planning

Breaches happen. Organizations that plan respond better—minimizing damage, recovering faster, maintaining stakeholder trust.

Incident Response Team

Designate a team with clear roles: Incident Commander (overall authority), Technical Lead (investigating technical aspects), Communications Lead (managing internal and external communication), Legal/Compliance (ensuring regulatory compliance), and Finance (tracking costs). During incidents, clear authority and roles prevent chaos.

Response Procedures

Documented procedures for incident response include: detection and reporting (how are incidents reported?), assessment (what's the scope and severity?), containment (stop the attack), investigation (understand what happened), remediation (fix the problem), recovery (restore normal operations), and post-incident review (what did we learn?). Run tabletop exercises simulating incidents before real breaches occur—these reveal gaps in planning.

Communication Plans

Plan communication to different audiences: staff (what they need to know to help respond), donors (especially if their data was breached), funders (compliance requirements), regulators (if legally required), and media (if public attention occurs). Different audiences need different messages at different times. Develop templates in advance so you're not writing communications while crisis unfolds.

Security Training and Awareness

Technical controls are insufficient—human behavior enables most breaches. Phishing (fake emails tricking staff into revealing credentials) is the most common attack vector. Comprehensive security training includes:

Initial Security Orientation

All new staff receive security training: password management, recognizing phishing, handling sensitive data, reporting suspicious activity. Training is mandatory, not optional.

Ongoing Awareness

Monthly or quarterly security awareness messages (newsletters, lunch-and-learns, tips) keep security top-of-mind. Current-event security awareness (educating about recent breaches and lessons) increases relevance.

Phishing Simulations

Send fake phishing emails to staff, tracking who clicks malicious links. Staff who fall for simulations receive additional training. This approach identifies vulnerable staff and reinforces awareness.

Compliance Monitoring

Compliance with internal policies and external regulations requires ongoing monitoring and verification.

Compliance Audits

Regular audits (quarterly or annually) assess compliance with security policies, data governance policies, and regulatory requirements. Internal audits are baseline; external audits by third parties provide independent assurance.

Security Scanning

Automated tools scan systems for vulnerabilities (known security weaknesses). Vulnerability scanning should be continuous—new vulnerabilities are discovered regularly. Prioritize fixing critical and high-severity vulnerabilities immediately.

Penetration Testing

Authorized attackers attempt to breach your systems. Penetration testing reveals vulnerabilities that automated scanning might miss and assesses whether controls actually prevent unauthorized access. Annual or biennial penetration testing by external firms provides valuable assessment.

Cyber Liability Insurance

No security program is perfect. Cyber liability insurance transfers some financial risk of breaches to insurers. Policies typically cover: notification costs, credit monitoring for affected individuals, forensic investigation, legal defense, and breach response costs. Insurance cannot eliminate need for security but provides financial backstop.

Lab: Enterprise Security Risk Assessment

Conduct a comprehensive security risk assessment for your organization's enterprise AI implementation: Identify 15-20 potential threats (data breaches, system failures, unauthorized access, etc.). For each, estimate likelihood and impact. Develop a risk prioritization matrix. For the highest-priority risks, develop specific mitigation strategies (technical controls, process changes, policy changes). Document your assessment and present mitigation plan to leadership.

Summary

Enterprise AI security protects the Confidentiality, Integrity, and Availability of sensitive organizational and stakeholder data. Threat modeling identifies what could go wrong. Risk assessment prioritizes mitigation efforts. Security controls frameworks organize comprehensive protection across identification, protection, detection, response, and recovery. Access management and encryption provide foundational controls. Vendor security assessment extends security to external partners. Incident response planning and communication prepare for breaches. Security training engages staff. Compliance monitoring ensures adherence. Cyber liability insurance provides financial backup. Organizations implementing these elements comprehensively protect themselves and stakeholders from AI-related security harms.

Ready to Master Enterprise AI for Your Nonprofit?

Enroll in CAGP Level 4 to deepen your skills in organizational-scale AI implementation, measurement, and strategy.

Explore CAGP Levels