Advisory on Risks associated with Frontier AI Models #AI

Executive Summary

Frontier Artificial Intelligence (AI) models are the most recent advanced AI models available. These models have demonstrated enhanced cybersecurity capabilities such as software analysis, vulnerability discovery, and security reasoning, at a level approaching or complementing cybersecurity practitioners. These frontier AI models can reportedly reduce the time taken to identify vulnerabilities and engineer exploits – cutting short the duration from months to hours. This will potentially reduce the time taken for developers to fix identified bugs. However, the same capability could also be misused by cyber threat actors to accelerate vulnerability exploitation and the development of malicious capabilities.

While there are no indications that such capabilities are being misused or abused at this point, this advisory outlines how organisations can plan ahead and strengthen their cybersecurity posture to guard against such risks.

Understanding Frontier AI Capabilities 

Frontier AI systems are capable of:

• Analysing large and complex codebases to identify security weaknesses 

• Detecting subtle logic flaws and insecure coding patterns 

• Supporting end-to-end vulnerability workflows, including: 

  • vulnerability identification 

  • exploitability reasoning in controlled environments 

  • patch suggestion and remediation guidance 

• Scaling security analysis across systems at a speed and breadth beyond traditional manual review 

What You Can Do To Protect Your Enterprise

Immediate Mitigation Measures:

• Patch all high-critical vulnerabilities

  • Ensure all critical and high-severity vulnerabilities on internet-facing systems have been remediated. These assets face the greatest exposure to automated attacks and present the highest risk of widespread impact if compromised.

• Implement Multi-Factor Authentication (MFA)

  • Ensure MFA is enabled on all administrative interfaces, gateways, and cloud management consoles. For systems that cannot support MFA, implement IP allow listing to restrict access to authorised sources only.

• Secure and disconnect development environments

  • Identify all internet-facing development, staging, and test environments, including their management interfaces. Implement strict access controls or, if this cannot be achieved, immediately disconnect these systems from internet access.

• Review cloud security configurations

  • Review and tighten all cloud security configurations and address all flagged issues or alerts. Prioritise publicly accessible resources, overly permissive security groups, and exposed management interfaces.

• Review and enforce least-privilege access

  • Review current user permissions to identify and remove unnecessary access rights, implement role-based access controls that align strictly with actual job functions, and revoke all dormant accounts and unused service accounts. Review system-to-system access by applying least privilege principles to service accounts and APIs to minimise potential attack surfaces.

• Enable DDoS Protection

  • Ensure DDoS protection services are enabled and correctly configured on all internet-facing assets. AI-powered tools can rapidly identify and exploit services with inadequate protection.

Longer Term Mitigation Measures:

Reducing attack paths and attack surfaces:

• Strengthen Perimeter Defence and System Hardening

  • Robust access controls and minimise attack surfaces by removing unnecessary internet-facing services, disabling unused ports and protocols, and securing cloud configurations. This directly counters AI-powered automated reconnaissance and vulnerability scanning that can rapidly identify and exploit exposed systems.

• Deploy Network Segmentation to Contain AI-Driven Attacks

  • Establish comprehensive network segmentation with micro-segmentation where feasible to limit lateral movement capabilities. AI-enhanced attacks can rapidly pivot through networks once initial access is gained, making containment through segmented architectures critical to preventing widespread compromise.

• Secure Supply Chain and Dependency Management

  • Implement rigorous supply chain security measures, including dependency scanning, controlled update policies, and vendor risk assessments. Cyber threat actors can leverage AI to accelerate the identification and exploitation of vulnerabilities in third-party components and dependencies, making proactive supply chain security essential to prevent compromises along the supply chain.

Monitor plausible attack paths:

• Implement Comprehensive Attack Path Monitoring

  • Deploy continuous monitoring across all critical attack vectors including network traffic, endpoint activities, and user behaviours. Focus surveillance on high-risk pathways such as privileged account usage, lateral movement patterns, and access to sensitive systems. AI-enhanced attacks can execute complex multi-stage operations rapidly, making real-time visibility across potential attack chains essential.

• Establish Anomaly Detection for AI-Driven Threats

  • Implement behavioural analytics and machine learning-based detection systems to identify subtle deviations from normal patterns that may indicate AI-powered attacks. Monitor for automated reconnaissance activities, unusual data access patterns, and accelerated attack progressions that exceed typical human-operated threat timelines.

Implement layered security (defence-in-depth):

• Deploy Defence-in-Depth Architecture

  • Implement multiple overlapping security layers across network, endpoint, application, and data tiers to ensure redundant protection. Ensure that the design of systems does not enable the compromise of overall defence posture through a single point of failure. This layered approach is critical against AI-enhanced attacks that can rapidly adapt and bypass individual security measures.

• Integrate Security Throughout Application Lifecycle

  • Embed automated security testing and validation at every stage of procurement, development, and deployment processes. Adopt zero-trust architecture principles and assume breach mentality by implementing continuous verification, least-privilege access, and runtime application security monitoring. This prevents AI from exploiting vulnerabilities introduced during software development or third-party integrations.

• Strengthen Identity and Access Management

  • Mandate MFA across all systems and implement rapid credential response capabilities including automated password resets and account lockouts. Deploy continuous authentication monitoring to detect and respond to compromised credentials within minutes rather than hours, as AI-powered attacks can escalate privileges and move laterally at machine speed once initial access is obtained.

Shorten patch cycles:

• Accelerate Critical Vulnerability Patching

  • Establish expedited patching cycles for high and critical CVE-rated vulnerabilities, treating them as urgent security incidents rather than routine maintenance tasks. Implement automated patch deployment systems where feasible to reduce manual delays and human error in the patching process.

• Minimise Vulnerability Exposure Windows

  • Reduce time-to-deploy for security updates by streamlining approval processes and pre-testing patch compatibility in isolated environments. AI-powered attacks can weaponise newly disclosed vulnerabilities within hours of publication, making rapid patch deployment critical to preventing mass exploitation.

Using AI to identify vulnerabilities:

• Deploy AI-Powered Vulnerability Detection Systems

  • Implement automated AI-assisted tools for continuous scanning and identification of system misconfigurations, weak credentials, and exploitable vulnerabilities across the entire IT infrastructure. These tools can match the speed and scale of AI-powered attacks by providing real-time vulnerability assessment capabilities that exceed traditional manual security audits.

• Maintain Comprehensive Asset Visibility and Inventory

  • Establish and continuously update accurate asset inventories that identify all systems, applications, and network components within the organisation. AI-enhanced vulnerability management requires complete visibility of the attack surface to ensure no critical systems are overlooked during vulnerability assessments and remediation efforts.

Conclusion

Frontier AI models represent a major advancement in enhancing cybersecurity capabilities but there are also risks involved. Organisations should take proactive steps to raise cyber hygience standards and strengthen overall cyber defence posture to protect themselves against risk of attacks from frontier AI models.