Full Report
Defenders can’t use AI in the same way attackers can, but there’s much they can do to unlock the potential of agentic cyber defence.
Analysis Summary
# Best Practices: Agentic Cyber Defence
## Overview
These practices address the inherent imbalance between AI-powered offensive cyber actions and defensive AI implementation. While attackers focus on technical success (crashes, exfiltration), defenders must navigate "political" and organizational constraints, ensuring that automated actions do not disrupt business operations. The goal is to move from manual intervention to controlled, agentic defense without compromising system availability.
## Key Recommendations
### Immediate Actions
1. **Identify Deterministic Automations:** Review current SOC operations to identify high-confidence, low-impact tasks that can be automated using SOAR (Security Orchestration, Automation and Response), such as forcing re-authentication for suspicious logins.
2. **Establish Risk Appetite:** Conduct a stakeholder meeting to define which business functions can tolerate automated remedial actions and which require human-in-the-loop (HITL) approval.
3. **Defensive Offensive Testing:** Deploy AI-enabled vulnerability scanning and penetration testing tools in pre-production/staging environments where the risk of business disruption is minimal.
### Short-term Improvements (1-3 months)
1. **Develop "Guardrail" Policies:** Create technical constraints for AI agents that prevent them from modifying critical infrastructure or firewall rules without secondary validation.
2. **Refine Success Metrics:** Move beyond "threats blocked" to "business uptime preserved during automated response" to align security goals with organizational mission.
3. **Pilot Shadow AI Discovery:** Use AI tools to identify "Shadow AI" usage within the organization to understand staff needs and security gaps.
### Long-term Strategy (3+ months)
1. **Autonomous Cyber Defence (ACD) Integration:** Gradually transition from scripted SOAR playbooks to agentic models that can reason through complex defensive tasks within predefined organizational constraints.
2. **Political-Technical Alignment:** Standardize the process for "Change Requests" to allow for pre-approved, automated emergency patching for high-criticality vulnerabilities.
3. **Architectural Resilience:** Redesign systems to be "agent-friendly," ensuring they provide the necessary telemetry for AI agents to make informed, safe defensive decisions.
## Implementation Guidance
### For Small Organizations
- Focus on "Cyber Essentials" basics first.
- Utilize SaaS-based security tools that incorporate "built-in" AI/ML for automated threat blocking rather than building custom agents.
### For Medium Organizations
- Implement SOAR for high-volume, low-risk alerts (e.g., phishing email analysis).
- Ensure a "Human-in-the-loop" model for any action that could affect network connectivity or user access.
### For Large Enterprises
- Adopt the **CETAS Autonomous Cyber Defence** framework to categorize AI appetite across different business units.
- Invest in sandboxing environments for testing agentic defensive tools before deployment in production.
## Configuration Examples
*While specific code was not provided in the text, the following is an actionable configuration strategy based on NCSC guidelines:*
**SOAR Policy Logic (Example):**
- **Trigger:** Suspicious travel/Impossible travel detected.
- **Action:**
- IF confidence score > 90% AND user is NOT C-suite: Trigger MFA reset (Automated).
- IF confidence score < 90%: Open ticket for analyst (Human-in-the-loop).
- IF user is C-suite: Alert SOC Lead immediately; do not lock account automatically.
## Compliance Alignment
- **NCSC Cyber Assessment Framework (CAF):** Aligning automated response with objective B2 (Resilient networks and systems).
- **NIST AI Risk Management Framework (AI RMF):** Focusing on the "Govern" and "Map" functions to manage organizational risk.
- **Cyber Essentials:** Ensuring automated patching processes meet the 14-day requirement.
## Common Pitfalls to Avoid
- **"Set and Forget" Mentality:** Assuming AI agents can operate without oversight; defenders, unlike attackers, are responsible for collateral damage (system downtime).
- **Ignoring the "Political" Problem:** Implementing technical blocks that interfere with business-critical functions without obtaining prior stakeholder buy-in.
- **Over-reliance on Non-Deterministic Tools:** Using AI that provides unpredictable results for high-stakes defensive actions.
## Resources
- **NCSC Guidance:** Artificial Intelligence Hub [www.ncsc.gov.uk/section/advice-guidance/all-topics/artificial-intelligence]
- **Framework:** CETAS Autonomous Cyber Defence (The Alan Turing Institute) [cetas.turing.ac.uk/publications/autonomous-cyber-defence-autonomous-agents]
- **Standards:** NCSC Cyber Assessment Framework (CAF) [www.ncsc.gov.uk/collection/caf]