Case study · AI security

Agents inside, attackers outside: AI cyber risk controls for a UK payments firm

A UK payments firm had AI agents in customer support, fraud operations and engineering, each built differently and each with its own access. At the same time, attacks on it were getting faster and more convincing. The goal was one way to build and govern agents, and controls that assume attackers use AI too, in a form the firm could show its regulators.

Case studyPayments · United KingdomSumanta Mukhopadhyay9 min read
#ai#security#governance
CASE STUDY · CASE STUDY · AI SECURITY20 Nov 2026 · 15:50 GST
At a glance
SectorPayments · United Kingdom
My roleFractional CTO and AI architect
EngagementAssessment, architecture and delivery leadership
ScopeAI agent inventory, agent platform, identity and tool access, monitoring, AI-enabled attack defences, operational resilience
What changed
  • Every AI agent inventoried with an owner, purpose and access
  • Agents built on one framework with shared controls
  • Agent tool access scoped, logged and revocable
  • Payment and access changes protected against impersonation
  • AI risk mapped to important business services and impact tolerances
Shared agent frameworkTool gatewayWorkload identityPolicy engineOpenTelemetry GenAI telemetrySecurity event pipelinePhishing-resistant authenticationOut-of-band verification

Anonymised composite. Client names, identifying details and confidential figures are withheld; public facts are cited in the references.

The situation

A UK payments and e-money firm had adopted AI quickly and unevenly. Customer support used an agent that could look up transactions and raise refunds. Fraud operations used one that summarised alerts and suggested account restrictions. Engineering used coding agents with access to repositories and deployment pipelines. Each was built by a different team, on a different framework, with its own service accounts and its own idea of logging.

At the same time, the attacks the firm saw were changing: more convincing phishing, impersonation of executives in calls and messages asking for urgent payment changes, and faster exploitation of newly published flaws. The firm already worked to the FCA's operational resilience rules, which require important business services to be identified and kept within impact tolerances3.

Why now. In November 2026 the Bank of England, PRA and FCA published a consultation on cyber risk and resilience management, CP17/26 and GC26/9, clarifying how firms should meet existing requirements, with responses due by 26 February 20272. Alongside it they published effective practices on AI-related cyber risk, which treat AI as creating new risks, amplifying existing ones and helping defenders, and which include reducing the attack surface against AI-enabled attacks1. The board wanted to know how the firm compared, before a supervisor asked.

What had to be true

  • Every agent known: owner, purpose, model, tools and data it can reach.
  • One way to build agents, so controls are written once.
  • Agent actions as accountable as staff actions: identity, least privilege, logs and a way to stop them.
  • Defences that assume attackers use AI, especially for payments and access changes.
  • Everything tied to important business services, so AI risk fits the resilience framework the firm already runs.

Architecture

Agents
Support agentlookups, refundsFraud agentsummaries, suggested restrictionsEngineering agentscode, pipelines
Agent platform
Shared frameworkone runtime, one config modelModel gatewayapproved models onlyPrompt and output filtersinjection, data leakage
Access
Tool gatewayscoped, logged, revocableWorkload identityper agent, short-livedPolicy enginelimits, approvals
Defence
Phishing-resistant sign-instaff and adminsOut-of-band verificationpayments and access changesFast patchingexposed services first
Assurance
Agent telemetryGenAI conventionsSecurity monitoringagent and human activityService mappingimpact tolerances
Agents go through one platform and one access layer; defences assume the attacker has AI too.

The data flow, end to end

  1. RequestA user or event triggers an agent through the shared framework
  2. IdentifyThe agent runs with its own short-lived identity, not a shared account
  3. Call toolsEvery tool call goes through the gateway and is checked against policy
  4. ApproveMoney movement and access changes need a person, verified out of band
  5. RecordPrompts, tool calls and outcomes are captured as telemetry
  6. DetectSecurity monitoring flags unusual agent behaviour alongside human activity
The tool gateway and the approval step carry most of the risk reduction.

The hard part: controlling agents the firm did not build the same way

The three agent estates could not be governed one by one; there was no shared place to put a control. The answer was to move them onto one framework over time, so identity, tool access, logging and a kill switch were properties of the platform rather than of each team's code. OpenTelemetry's work on agent observability draws the same distinction between agent applications and the frameworks they are built on, and argues for standard telemetry at the framework level so every application inherits it5.

Diagram of one AI agent framework at the top with arrows to several AI agent applications, each containing its own model and tools
One framework, many agent applications: put controls and telemetry in the framework and every agent inherits them. Source: OpenTelemetry blog, Guangya Liu and Sujay Solomon · CC BY 4.0

The fastest risk reduction came before any migration: an inventory, then removing tools agents did not need. Some had write access they had never used.

The second half of the hard part was outside the firm. Impersonation attacks target the same processes agents touch: refunds, payee changes, admin access. Those changes now need a person to approve them through a separate, verified channel, never on the strength of a message, call or agent request alone. The pattern follows the deepfake-resistant onboarding and staff access hardening case studies, and the agent platform builds on the enterprise MCP agent platform.

How it ran

Risk and compliance
  1. Map AI to important business services
  2. Compare against effective practices
  3. Consultation response
Platform and security
  1. Agent inventory
  2. Tool gateway and identities
  3. Telemetry and monitoring
Operations
  1. Approval steps for payments and access
  2. Staff training on impersonation
  3. Agent incident playbooks
Risk framed it, platform and security built it, operations made the human checks work.
Phase 1Inventory
  • Find every agent
  • Owners and access
  • Remove unused tools
Phase 2Contain
  • Tool gateway
  • Per-agent identity
  • Out-of-band approvals
Phase 3Standardise
  • Shared framework
  • Telemetry
  • Kill switches
Phase 4Evidence
  • Service mapping
  • Scenario tests
  • Board and regulator view
Know what you have, contain it, standardise it, then prove it.

Key decisions and trade-offs

DecisionAlternativeWhy
One agent framework for the firmEach team chooses its ownControls built once and applied everywhere
All tool calls through a gatewayAgents call APIs directlyOne place to scope, log and revoke access
Per-agent, short-lived identitiesShared service accountsActions traceable to an agent; stolen tokens expire
Human approval for money and access changesTrust agent or message aloneImpersonation and injection target exactly these steps
Map to existing resilience frameworkA separate AI risk registerAI risk sits where the board already looks

What I'd do again, and differently

Again: start with the inventory and the tool gateway, which reduced risk before any larger change, and treat the regulators' effective practices as a checklist rather than waiting for final guidance. Differently: involve the contact centre and finance teams from day one. The impersonation attempts reached them first, and their ideas made the approval steps practical. The NCSC's Cyber Assessment Framework, which the consultation aligns with4, was a useful common language between engineering and risk. See vendor admin access as perimeter and the tech news for 20 November.

If you run a regulated UK business, my work with UK companies explains how I approach it, or analyse your challenge and bring the result to a call.

References
  1. 1
  2. 2
  3. 3
    PS21/3: Building operational resilienceFinancial Conduct Authority · Mar 2021
  4. 4
    Cyber Assessment FrameworkNational Cyber Security Centre
  5. 5
    AI Agent Observability: Evolving Standards and Best PracticesOpenTelemetry blog (Guangya Liu, Sujay Solomon) · 2025
FAQ

Common questions

What did UK regulators publish on AI cyber risk?

In November 2026 the Bank of England, PRA and FCA published a consultation on cyber risk and resilience management, CP17/26 and GC26/9, and separate effective practices describing how firms are managing AI-related cyber risk, including reducing their attack surface against AI-enabled attacks.

Are effective practices mandatory?

No. They describe practices regulators consider effective. In practice they become a reference point in supervisory discussions, so firms should know how their own controls compare.

Why standardise on one agent framework?

Controls such as identity, tool access, logging and kill switches can then be built once and applied to every agent, instead of being re-implemented, or forgotten, in each team's own build.

Keep reading

Related

Open to advisory, fractional CTO & select builds

Let's talk about what you're building.

The fastest route is a free 30-minute call. Prefer to write? Send a message and I'll reply personally within two business days.

Start a conversation

How would you like to reach me?

Book a 30-min callFastest. Pick any open slot on my calendar.→ Write in GmailOpens a pre-filled draft in your browser.→ Use my email appOutlook, Apple Mail or your default client.→

I reply personally within two business days. Time zones from the Americas and Europe to the Middle East and Asia-Pacific.