A 26-person AI agency in Philadelphia got a call that changed their approach to security forever. Their largest client — a financial services company representing 28% of revenue — informed them that they had failed a third-party security assessment. The assessment, conducted as part of the client's own regulatory compliance, revealed that the agency had unencrypted client data on developer laptops, no formal access control policies, shared passwords for cloud accounts, and no incident response plan. The client gave the agency 60 days to remediate all findings or face contract termination. The remediation cost $85,000 in consultant fees, tool purchases, and lost productivity. But the real cost was the three months of management distraction and the lingering trust damage with their most important client.
The agency's founder admitted: "We had great security practices for the AI systems we built for clients. We had terrible security practices for our own organization. Nobody thought to audit ourselves."
This is a common pattern in AI agencies. Engineers who are meticulous about securing the systems they build for clients often work in agencies with minimal internal security practices. And as AI agencies handle increasingly sensitive data — healthcare records, financial data, personal information, proprietary business data — the risk of a security incident grows.
Why AI Agencies Need Security Audits
Client Data Exposure
AI agencies routinely handle sensitive client data — training datasets, customer records, business metrics, proprietary algorithms. A data breach that exposes client data can result in legal liability, contract termination, regulatory penalties, and reputational damage.
Enterprise Client Requirements
Enterprise clients increasingly require their vendors to demonstrate security compliance. SOC 2, ISO 27001, and industry-specific frameworks (HITRUST for healthcare, PCI DSS for payment data) are becoming table stakes for winning and retaining enterprise AI contracts.
Regulatory Obligations
Data privacy regulations (GDPR, CCPA, state privacy laws) impose security requirements on any organization that processes personal data. Non-compliance can result in significant fines — up to 4% of global revenue under GDPR.
Intellectual Property Protection
Your agency's proprietary frameworks, methodologies, and trained models are valuable intellectual property. A security breach that exposes your IP undermines your competitive advantage.
Supply Chain Attacks
As an agency that builds AI systems deployed in client environments, you are part of your clients' supply chain. A compromise of your systems could be used as a vector to attack your clients. Supply chain security is under intense scrutiny following high-profile incidents.
The Security Audit Framework
A comprehensive security audit for an AI agency should cover seven domains.
Domain 1 — Access Control and Identity Management
What to audit:
- User account management: How are accounts created, modified, and deactivated? Are there orphan accounts for former employees or contractors?
- Authentication: Is multi-factor authentication (MFA) enforced on all business-critical systems? Which systems lack MFA?
- Password policies: Are strong passwords required? Is a password manager provided and mandated?
- Access reviews: How often are access permissions reviewed? Can you demonstrate that access is based on the principle of least privilege?
- Privileged access: Who has administrative access to cloud accounts, production systems, and client environments? Is privileged access monitored and logged?
- Shared accounts: Are any accounts shared between team members? Shared accounts eliminate accountability and should be eliminated.
Common findings in AI agencies:
- MFA not enabled on cloud provider console (critical risk)
- Former contractors still have active access to client repositories (critical risk)
- Shared root/admin credentials for cloud accounts (high risk)
- No formal process for revoking access when team members leave (high risk)
- Over-permissioned accounts — everyone has admin access to everything (medium risk)
Remediation priorities:
- Enable MFA on all cloud accounts, email, code repositories, and client systems immediately
- Conduct an access review and remove all unnecessary permissions
- Eliminate shared accounts and implement individual accounts with role-based access
- Establish an offboarding checklist that includes access revocation for all systems
Domain 2 — Data Protection
What to audit:
- Data classification: Do you classify data by sensitivity level? Do team members know how to handle each classification level?
- Encryption at rest: Is data encrypted on laptops, cloud storage, databases, and backups?
- Encryption in transit: Is all data transmitted over encrypted connections (TLS/HTTPS)?
- Data handling procedures: How is client data received, stored, processed, and returned or destroyed? Is there a documented procedure?
- Data retention: How long is client data retained after project completion? Is there a defined retention policy?
- Data disposal: How is client data destroyed when no longer needed? Is disposal documented?
- Laptop/device security: Are laptops encrypted? Is remote wipe capability enabled? What happens if a laptop is lost or stolen?
- Data backups: Are backups encrypted? Are they stored separately from primary data? Are they tested for recoverability?
Common findings in AI agencies:
- Client training data stored unencrypted on developer laptops (critical risk)
- No data retention or disposal policy — client data from three-year-old projects still on shared drives (high risk)
- No laptop encryption enforced (high risk)
- Client data shared via unencrypted email or personal cloud storage accounts (high risk)
- No formal data handling procedures — each team member manages data differently (medium risk)
Remediation priorities:
- Enforce full-disk encryption on all company and personal devices used for work
- Implement a data handling policy that specifies how client data is received, stored, processed, and disposed
- Delete client data from completed projects per a defined retention schedule
- Provide secure file transfer mechanisms and prohibit client data in email or personal cloud storage
Domain 3 — Network and Infrastructure Security
What to audit:
- Cloud infrastructure: Are cloud accounts configured securely? Are security groups, network ACLs, and IAM policies properly configured?
- Network segmentation: Is client data isolated from other workloads? Are development, staging, and production environments separated?
- Vulnerability management: Are systems patched regularly? Are there known vulnerabilities in your infrastructure?
- Firewall and perimeter: Are unnecessary ports and services exposed? Are management interfaces protected?
- Remote access: How do team members access cloud and client systems remotely? Is VPN or zero-trust network access used?
- Monitoring and logging: Are security events logged? Are logs monitored for suspicious activity? How long are logs retained?
Common findings in AI agencies:
- Cloud infrastructure with default security configurations (medium to high risk depending on exposure)
- No network segmentation — all projects share the same infrastructure (medium risk)
- Inconsistent patching — some systems months behind on security updates (medium to high risk)
- No centralized logging or monitoring — security events would go undetected (high risk)
- SSH keys or API credentials stored in code repositories (critical risk)
Remediation priorities:
- Conduct a cloud security configuration review against your provider's security best practices (AWS Well-Architected, Google Cloud Security Command Center, Azure Security Benchmark)
- Implement centralized logging for all cloud resources and critical systems
- Establish a patching cadence (critical patches within 48 hours, high within 7 days, medium within 30 days)
- Scan all code repositories for exposed credentials and rotate any that are found
Domain 4 — Application Security
What to audit:
- Secure development practices: Do developers follow secure coding guidelines? Is there a secure development lifecycle?
- Code review: Are all code changes reviewed by another developer before deployment?
- Dependency management: Are third-party libraries and dependencies monitored for known vulnerabilities?
- Secrets management: How are API keys, database credentials, and other secrets managed? Are they stored securely or hardcoded?
- AI/ML-specific security: Are model artifacts protected? Are training pipelines secured? Is there protection against adversarial inputs, model extraction, or data poisoning?
- Client-facing applications: Are web applications and APIs tested for common vulnerabilities (OWASP Top 10)?
Common findings in AI agencies:
- API keys and credentials hardcoded in source code or configuration files (critical risk)
- No dependency scanning — known vulnerabilities in third-party libraries (high risk)
- No secure development guidelines — security left to individual developer judgment (medium risk)
- Model artifacts (trained weights, preprocessing pipelines) not protected against unauthorized access (medium risk)
- No testing for AI-specific attacks like adversarial inputs or prompt injection (medium risk)
Remediation priorities:
- Implement a secrets management solution (HashiCorp Vault, AWS Secrets Manager, or similar)
- Enable automated dependency scanning in your CI/CD pipeline
- Establish secure coding guidelines with AI/ML-specific considerations
- Protect model artifacts with appropriate access controls and integrity verification
Domain 5 — Physical Security
What to audit:
- Office security: If you have a physical office, how is access controlled? Are servers or sensitive equipment secured?
- Remote work security: What requirements exist for home office security? Are team members trained on physical security for devices?
- Device management: How are company devices tracked? What happens when a device is lost, stolen, or damaged?
- Visitor access: How are visitors managed in the office? Do they have access to work areas where sensitive information is visible?
Common findings in AI agencies (especially those with remote teams):
- No mobile device management — lost devices cannot be remotely wiped (high risk)
- No policy for working in public spaces — team members working on client data in coffee shops and airports (medium risk)
- No clean desk policy — sensitive information visible on desks or screens (low to medium risk)
Domain 6 — Incident Response
What to audit:
- Incident response plan: Do you have a documented plan for responding to security incidents?
- Roles and responsibilities: Does everyone know their role in an incident? Is there a defined incident commander?
- Communication plan: How will you communicate with affected parties — clients, employees, regulators, law enforcement?
- Detection capability: Can you detect a security incident when it occurs? What monitoring and alerting is in place?
- Recovery procedures: Can you recover from a security incident? Are backups tested? Is there a business continuity plan?
- Post-incident review: Do you have a process for learning from security incidents?
Common findings in AI agencies:
- No incident response plan (critical risk)
- No defined roles or communication procedures for security incidents (high risk)
- No tested backup or recovery procedures (high risk)
- No security monitoring or alerting — incidents would go undetected (high risk)
Remediation priorities:
- Create a basic incident response plan covering detection, containment, eradication, recovery, and lessons learned
- Define roles and responsibilities and ensure key personnel know their assignments
- Test backup recovery procedures quarterly
- Implement basic monitoring and alerting for critical systems
Domain 7 — Third-Party and Supply Chain Security
What to audit:
- Vendor assessment: Have you assessed the security posture of your critical vendors (cloud providers, SaaS tools, subcontractors)?
- Subcontractor security: Do subcontractors with access to client data meet your security requirements?
- Data sharing: What data is shared with third parties? Is sharing governed by appropriate agreements?
- Open-source components: Are open-source libraries and models assessed for security before use?
Common findings in AI agencies:
- No vendor security assessment process (medium risk)
- Subcontractors with access to client data but no security requirements in their contracts (high risk)
- Pre-trained models downloaded from public repositories without integrity verification (medium risk)
Conducting the Audit
Internal vs. External Audit
Internal audit: Conducted by your own team using the framework above. Cost-effective and can be done frequently. Appropriate for ongoing monitoring and improvement.
External audit: Conducted by a third-party security firm. More objective, more thorough, and carries more credibility with clients. Appropriate annually or when pursuing compliance certifications.
Recommended approach: Conduct an internal audit quarterly using the framework above. Engage an external firm annually for a comprehensive assessment. The internal audits keep you on track; the external audit validates your posture and identifies blind spots.
Audit Process
Step 1 — Scope and plan (1-2 days): Define which domains will be covered, gather documentation, and schedule interviews with key personnel.
Step 2 — Assessment (3-5 days for internal, 1-2 weeks for external): Review documentation, interview team members, examine configurations, test controls, and identify vulnerabilities.
Step 3 — Report (2-3 days): Document findings with risk ratings (critical, high, medium, low), affected systems, and recommended remediations.
Step 4 — Remediation planning (1-2 days): Prioritize findings by risk and develop a remediation timeline. Critical findings should be addressed within 48 hours. High findings within 30 days. Medium within 90 days.
Step 5 — Remediation execution (ongoing): Implement fixes according to the plan. Track progress and validate that remediations are effective.
Step 6 — Verification (1-2 days): Re-test remediated findings to confirm they are resolved.
Building a Security-First Culture
An audit is a point-in-time assessment. Sustaining good security requires embedding security into your agency's culture and operations.
Security awareness training: Train all team members on security basics — phishing recognition, password hygiene, data handling, physical security, incident reporting. Repeat annually and supplement with targeted training when threats or practices change.
Security champions: Designate a security champion on each project team — an engineer who takes ownership of security considerations for that project. Security champions receive additional training and serve as the first line of defense.
Secure by default: Configure your systems and tools with security defaults — MFA enabled, encryption on, minimal permissions. Make the secure option the easy option.
Blameless incident culture: When security incidents or near-misses occur, focus on learning rather than blame. If people are afraid to report security issues, problems will be hidden until they become crises.
Your Next Step
Schedule your first internal security audit within the next 30 days. Start with the two highest-risk domains — access control and data protection. Assign an owner for each domain, give them the audit questions from this framework, and have them assess your current state. Document findings, prioritize by risk, and address critical findings immediately. If you do not have MFA enabled on all cloud accounts and email today, stop reading and enable it now — that single action eliminates a large percentage of account compromise risk. Then work through the remaining domains over the next quarter. A security audit is not a one-time project — it is the beginning of an ongoing security practice that protects your agency, your clients, and your reputation.