Bank of England, together with the PRA and FCA, publishes effective practices for managing AI cyber risk and resilience
Effective practices: Cyber risk and resilience management relating to artificial intelligence (AI)
The Bank of England, the Prudential Regulation Authority (PRA) and the Financial Conduct Authority (FCA) jointly published a document setting out effective practices for financial institutions and financial market infrastructures to manage AI-related cyber risk and resilience. The document judges that AI is significantly reducing the cost, skills and time for threat actors to discover and exploit vulnerabilities, with the exploitation cycle potentially shortened from weeks to hours or days, and that firms need to shift from traditional vulnerability management to continuous exposure management, with boards and executives accountable for cyber risk as a core business risk. The practices cover identifying exposure surfaces such as AI systems and third-party dependencies, building harnesses around advanced AI systems to obtain verifiable defensive outputs, using phishing-resistant multi-factor authentication to guard against deepfakes, conducting threat modeling that includes adversarial prompts and jailbreaks, developing response and recovery plans for model poisoning and prompt injection, and retaining meaningful human oversight. The document cites UK AISI evaluations saying AI performance in some domains doubles every eight months and that guardrails are improving but vulnerabilities remain; these practices are not formal requirements by regulators, and firms may adopt them as appropriate to their size and risk profile.