> For the complete documentation index, see [llms.txt](https://docs.coherent.global/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.coherent.global/trust-center/coherent-trust-center/frontier-ai-supplier-resilience-faq.md). # Frontier AI Supplier Resilience - FAQ Frontier AI refers to the latest generation of advanced artificial intelligence technologies, characterised by increasingly sophisticated capabilities in areas such as reasoning, software development, data analysis and autonomous task execution. While these advances have the potential to drive significant innovation, productivity and business value, they also introduce new cyber security, operational resilience and governance challenges. \ As the threat landscape continues to evolve, it is important that we assess how critical services, supporting technologies and control environments can adapt to emerging AI-enabled risks. This includes considerations across cyber defence, vulnerability management, incident detection and response, third-party risk management, and service continuity.
How are you adapting your cyber defence, monitoring, and vulnerability management capabilities to address rapidly advancing AI-enabled cyber threats? AI is increasing the speed, scale, and sophistication of cyberattacks, but it does not replace established security fundamentals. Our priority is to strengthen identity, data, endpoint, cloud, application, and vulnerability controls; add protections specifically for AI systems; and use automation responsibly to improve detection and response speed. * **Enhanced preventive controls:** Strengthening identity security with least privilege, privileged-access management, zero-trust controls, email security, endpoint protection, and application-layer safeguards. * **Using AI defensively :** with human oversight, Applying approved AI tools to enrich alerts, correlate events, summarise incidents, identify anomalies, and accelerate investigations, while retaining human validation for material decisions. * **AI-Proxy :** Coherent implements a secure, enterprise-grade HTTP reverse proxy for third-party AI APIs. It provides JWT authentication, transparent proxying, centralised credential management, usage tracking, and endpoint governance. Customer data is never used to train our models or those of any third party. All third-party AI traffic passes through our governed AI proxy, with no cross-tenant data exposure. * **Accelerating vulnerability management:** Prioritising vulnerabilities by exploitability, asset criticality, external exposure and business Impact. Internet-facing and actively exploited vulnerabilities receive remediation as defined in our Information Security Policy. * **Strengthening governance:** Maintaining an inventory of approved AI use cases, performing security and privacy assessments, monitoring third-party providers, and enforcing policies for sensitive data, and responsible AI use.
How quickly can you identify, assess, and remediate critical vulnerabilities? We apply risk-based vulnerability management SLAs based on severity, active exploitation, internet exposure, asset criticality, and business impact. We continuously scan our critical assets and prioritise the resolution as agreed in our Information Security Policy. Critical vulnerabilities are usually patched within 5 business days. If patching cannot be completed in this timeframe, compensating controls must be put in place within the same timeframes.
How do you assess cyber resilience risks introduced through your own third parties, open-source dependencies and AI providers? We assess cyber resilience risk from third parties, open-source dependencies, and AI providers through a structured vendor risk programme: * **Third parties:** Vendors are risk-tiered based on criticality and data access. Higher-risk vendors (e.g., cloud infrastructure and hosting providers) undergo security due diligence before onboarding, have contractual security and confidentiality obligations, and are subject to ongoing monitoring of their security posture, certifications, and incident history. * **Open-source dependencies:** Software composition analysis (SCA) and dependency scanning run as part of our CI/CD pipeline, with an SBOM maintained per service. Known-vulnerable dependencies are remediated under the severity-based SLAs. * **AI providers:** Third-party AI/LLM providers are vetted for their own security and data-handling practices before approval, and all AI API traffic is routed through our governed AI-Proxy - centralized credential management, usage tracking, and no cross-tenant data exposure.
What testing have you performed against AI-enabled attack techniques? We test our systems against AI-enabled attack techniques through our established QA and security-testing lifecycle. Our programme includes: **Secure QA and code review:** Standard functional, integration, negative, and regression testing is performed for application and security controls, including AI-enabled features and integrations. **Automated security testing:** We use Static Application Security Testing (SAST) and Dynamic Application Security Testing (DAST) to identify vulnerabilities in source code and running applications. Where applicable, dependency scanning, secret scanning, and software composition analysis are also performed. **Manual penetration testing:** Risk-based manual penetration testing is conducted to assess application, Integrated APIs, authentication, authorisation, data-protection, and the weaknesses that automated tools may not detect. **AI-specific abuse-case testing:** Where applicable, testing covers prompt injection, sensitive-data disclosure, insecure output handling, unauthorised access, cross-user or cross-tenant data exposure, manipulation of AI-generated output, and misuse of connected tools or APIs.
How would critical services continue to operate during a large-scale AI-enabled cyber incident? Our objective is to preserve the most critical services safely, operate in a controlled degraded mode, prevent the incident from spreading, and restore normal operations from verified clean environments. AI may accelerate an attack, but effective continuity still depends on segmentation, redundancy, trusted access, tested manual procedures, protected backups, and disciplined crisis management. Coherent maintains secure backups, BCP and DR plan implemented and tested periodically.
How are you using AI as part of your defence? **Threat detection:** We use AI to identify anomalous behaviour, emerging attack patterns, malicious emails, and suspicious identity, endpoint, and cloud activity. **Security operations:** AI helps correlate and prioritise alerts, enrich investigations, summarise incidents, and recommend response actions, reducing analyst workload and response times. **Vulnerability management:** AI supports risk-based prioritisation using exploitability, threat intelligence, asset criticality, exposure, and potential business impact. **Controlled use and oversight:** AI operates with human validation, least-privilege access, protected data, comprehensive logging.
How are you developing specific plans to mitigate the risk posed by the use of AI vulnerability identification LLMs, including the managing of scenarios involving automated exploitation of vulnerabilities and critical vulnerabilities (zero-day)? AI tools that can identify vulnerabilities shorten the time between a flaw existing and it being exploited, and increase the chance of attackers finding novel (zero-day) issues. We assume we cannot prevent every discovery, so our plan focuses on finding flaws first, fixing the riskiest ones fastest, detecting exploitation early, and containing the impact. * **Finding flaws before attackers do:** Secret scanning and static analysis, including dependency checks against known CVEs, run as part of our development lifecycle. We also commission independent manual penetration testing, which now includes AI-enabled attack techniques. * **Fixing the riskiest ones fastest:** Our vulnerability SLAs weight active exploitation and internet exposure. This matters even more as AI shortens attacker timelines: actively exploited and internet-facing critical vulnerabilities follow an emergency path, and where immediate patching is not possible we apply compensating controls such as restricting access, disabling the affected feature, or isolating services. * **Detecting exploitation:** Zero-day exploitation cannot be prevented by patching, so we rely on behavioural detection rather than signatures. We use managed threat detection across our cloud environment to flag anomalous API activity, credential misuse and unusual network behaviour, alongside the SOC operations. Least-privilege access and segmentation limit the blast radius if a flaw is exploited before it is known. * **Responding to zero-days:** Suspected critical or zero-day findings are escalated immediately to the Security and Engineering teams for validation and severity assessment. Affected systems are identified, temporary controls applied (disabling the affected feature, restricting access, isolating services, rotating potentially exposed credentials), and relevant logs preserved while permanent remediation is developed under our emergency handling process. Affected customers are notified in line with our contractual obligations. * **Governing our own AI tooling:** Approved LLM services access only authorised repositories, through restricted, least-privilege, read-only permissions. AI development tools operate under developer control; every change goes through human code review and our standard CI/CD gates, and no AI tool has autonomous merge, deploy, or production access. All repository and LLM-access activity is logged and monitored, and we can immediately suspend an LLM integration and revoke its credentials if needed. Third-party AI API traffic passes through our AI-Proxy, and we test AI-specific abuse cases including prompt injection, unauthorised repository traversal, and sensitive-data exfiltration.