bankofengland.co.uk web signal

PRA, FCA and Bank of England Issue AI Cyber Defence Guide

TL;DR

  • The PRA, FCA and Bank of England jointly published non-binding effective practices on AI cyber risk and resilience on 19 November 2026.
  • Named AI-specific threats include prompt injections, model and data poisoning, jailbreaking, adversarial inputs, autonomous AI agent attacks, and deepfakes.
  • The paper pushes firms from traditional vulnerability management toward continuous exposure management, with 'harness engineering' as a defensive discipline.

The three UK financial regulators have put out a joint document telling banks, insurers and market infrastructures how to treat AI on both sides of their cyber perimeter. The effective practices paper, published on 19 November 2026 by the PRA, the FCA and the Bank of England, is explicitly not a rulebook. The authorities describe it as complementary to formal supervisory expectations rather than a set of mandatory requirements.

The framing is blunt. "cyber risk is no longer a purely technical issue. It is a core business risk and leadership responsibility," the document says. The threats it names include prompt injections, model poisoning, data poisoning, jailbreaking, adversarial input attacks, autonomous AI agent attacks, deepfakes and sophisticated social engineering.

Speed is the central worry. "Advanced AI systems can analyse source code, identify vulnerabilities and discover attack paths at great speed... allowing threat actors to exploit vulnerabilities within hours or days rather than weeks or months," the paper says. Its answer is a shift from traditional vulnerability management, "where discovery has historically been the main bottleneck, towards continuous exposure management."

The paper also introduces "harness engineering" as a defensive discipline, defining a harness as "the collection of tools, workflows, controls, data and operating environments that sit around advanced AI systems and shape how they are used." It argues that "the number of vulnerabilities a firm discovers and remediates may depend as much on the harness as on the underlying AI." Human oversight, the authorities write, must be "meaningful, risk-based and capable of changing outcomes."

A separate consultation paper, CP17/26 and GC26/9, on technology-neutral cyber resilience was published the same day and closes for industry feedback on 26 February 2027; the AI document is designed to sit alongside it. The guidance arrives against a run of regulation coverage and fresh cybersecurity stories our tracker has logged in recent weeks, including the White House cyber director transition announced earlier the same day.