Purpose of the Template
The Financial Conduct Authority (FCA) has highlighted a significant issue: frontier AI models can expose cybersecurity vulnerabilities faster than your current controls can manage. This creates a risky period where your firm is aware of threats but hasn't yet addressed them.
This template guides you in conducting an AI-specific cyber risk assessment. It maps the AI tools your firm uses (or plans to use) against potential vulnerability scenarios. It's tailored for compliance officers at FCA-regulated entities who need to document how AI intersects with your cybersecurity measures and show supervisors that you're managing the dual-use nature of these technologies.
Use this to pinpoint where AI enhances threat detection and where it might accelerate threat exploitation, then assign responsibility for addressing these gaps.
Prerequisites
Before using this template, ensure you have:
- Current AI tool inventory: List every AI-powered system your firm uses for cybersecurity, fraud detection, customer service, trading algorithms, or operational automation.
- Recent penetration test or vulnerability scan results: Establish baseline data on known weaknesses.
- Incident response plan ownership matrix: Identify who responds when AI-assisted attacks occur.
- Third-party AI vendor contracts: Focus on SLAs around model updates and security patching.
- Your ISO/IEC 27001 Statement of Applicability or equivalent control framework: Integrate this assessment into your ISMS.
If you lack a current vulnerability baseline, pause and conduct that assessment first. This template assumes you know your existing weaknesses; the goal is to evaluate how AI changes your risk profile.
The Template
Copy this structure into your documentation system. Complete one assessment per AI system or capability cluster.
AI Cyber Risk Assessment Record
Assessment Date: [Date]
Assessed By: [Name, Title]
Review Frequency: [Quarterly recommended for frontier AI; annually for mature, stable AI tools]
Section 1: AI System Identification
- System Name/Description:
- Primary Function: [e.g., vulnerability scanning, phishing detection, access behavior analysis]
- Model Type: [Rule-based, machine learning, large language model, computer vision, etc.]
- Deployment Status: [Production / Pilot / Planned]
- Vendor/Developer: [Internal / Third-party name]
- Data Access Level: [What systems and data can this AI tool access?]
Section 2: Defensive Capability Assessment
How does this AI system improve your cybersecurity posture?
- Vulnerabilities It Identifies: [List specific vulnerability classes this tool detects]
- Speed Advantage: [How much faster does AI detection occur vs. manual methods?]
- Coverage Expansion: [What attack surfaces can you now monitor that you couldn't before?]
- Integration Points: [Which SIEM, SOAR, or monitoring platforms does this feed?]
Section 3: Threat Amplification Assessment
If an attacker had access to this same AI capability, how could they exploit it?
- Reconnaissance Acceleration: [Could this tool help attackers map your environment faster?]
- Vulnerability Exploitation: [Could findings be weaponized before you patch?]
- Social Engineering Enhancement: [Could AI-generated content bypass your awareness training?]
- Evasion Techniques: [Could attackers use similar AI to avoid your detection systems?]
Section 4: Control Gap Analysis
For each threat amplification scenario above, document your current control status:
| Amplification Scenario | Existing Control | Control Adequacy (Strong/Adequate/Weak/None) | Gap Description | Owner | Target Close Date |
|---|---|---|---|---|---|
| [Example: AI-generated phishing emails bypass email filters] | [Email security gateway with static rules] | [Weak] | [Filters don't analyze content patterns that LLMs generate] | [CISO] | [Q2 2024] |
Section 5: Vendor Risk (Third-Party AI Only)
- Model Update Frequency: [How often does the vendor release new versions?]
- Security Patching SLA: [Vendor commitment to fix vulnerabilities]
- Data Handling: [Where is training data stored? Is your firm's data used for model improvement?]
- Right to Audit: [Can you review vendor security controls?]
- Exit Strategy: [How quickly can you decommission this tool if it becomes compromised?]
Section 6: Regulatory Alignment
- FCA Principle Mapping: [Which FCA Principles for Businesses does this system support or risk violating?]
- Customer Impact: [Could malicious AI use harm customer data, market integrity, or financial stability?]
- Notification Triggers: [At what threshold would an AI-related incident require FCA notification?]
Section 7: Monitoring and Validation
- Performance Metrics: [How do you measure if this AI system is working as intended?]
- Drift Detection: [How do you identify when model behavior changes unexpectedly?]
- Human Oversight Checkpoints: [Where do humans review AI decisions before action?]
- Audit Trail: [Can you reconstruct why the AI made specific decisions?]
Customization Options
For smaller firms: Combine Sections 3 and 4 into a single "Risk vs. Control" table. Focus on documenting obvious dual-use risks and your mitigation plan.
For firms using frontier AI models: Add a subsection under Section 3 called "Emergent Capabilities." Document how you'll test for new capabilities that weren't present when you initially deployed the system.
For firms subject to multiple regulators: Duplicate Section 6 for each regulatory framework. An FCA-regulated firm with EU operations should map to both FCA Principles and the EU AI Act risk categories.
For third-party-heavy environments: Expand Section 5 into a full vendor risk assessment. Include fourth-party risk (your vendor's AI subprocessors) and contractual provisions around model explainability.
Integration with existing frameworks: This template should feed your ISO/IEC 27001 risk treatment plan (Clause 6.1.3), your NIST Cybersecurity Framework Identify and Protect functions, or your internal risk register. Map findings to control numbers in your existing ISMS.
Validation Steps
After completing your first assessment:
Cross-check against your threat model: Ensure every AI system appears in your threat landscape documentation. If you're modeling adversary tactics (e.g., MITRE ATT&CK), update those matrices to reflect AI-assisted attack scenarios.
Test control effectiveness: For each control marked "Strong" or "Adequate," schedule a validation exercise. If you claim your email filters catch AI-generated phishing, run a red team exercise with LLM-crafted emails.
Review with your Computer Security Incident Response Team: Walk through Section 3's threat amplification scenarios. Can your CSIRT detect and contain these AI-assisted attacks? Update your playbooks accordingly.
Submit to your risk committee: This assessment should inform board-level discussions about AI adoption. The FCA expects senior management accountability for technology risk; your completed template is evidence of that governance.
Schedule reassessment triggers: Set calendar reminders for your review frequency, but also define event-based triggers: major model updates, new AI tool deployments, post-incident reviews, or regulatory guidance changes.
If you identify control gaps you can't close within your target timeline, document the residual risk acceptance in your risk register and escalate to your risk committee. The FCA's concern isn't that AI creates risks; it's that firms adopt AI faster than they can govern it. A documented gap with a mitigation plan is better than an undocumented surprise.
