In partnership with
Good morning. Here’s what moved in AI risk, regulation and accountability across financial services this week. Two supervisors on opposite sides of the world landed on the same idea this week: judge an AI agent by what it can do when it fails, and make sure one named person answers for it. Meanwhile, India and Korea showed why that matters, with AI-assisted attackers already working through financial firms' defences.
In this week's Accountability Record:
🇺🇸 🟡 The OCC hands examiners a four-tier playbook for bank AI
🇭🇰 🟡 Hong Kong wants one named owner for every AI agent
🇮🇳 🔴 India issues a critical alert on AI-driven attacks against fintechs and lenders
🇰🇷 🟡 Korea's breach count more than doubles, and CEOs are in scope
Frontier models perform better in Attio’s agent harness.
We gave GPT-6 Astra the same job twice: analyze 100 sales calls and draft the follow-ups, once in Ask Attio and once in ChatGPT.
Ask Attio finished in 4 minutes 49 seconds. ChatGPT took 26 minutes 11 seconds, about 5.4 times as long.
The most ambitious GTM work requires more than frontier intelligence. It requires architecture that can put that intelligence to work at scale.
🇺🇸 US | Banking | Governance, model risk 🟡
The OCC hands examiners a four-tier playbook for bank AI
The gap: The US national bank supervisor now has a written method for examining AI, and it rates use cases by the harm a failure causes, not by how clever the model is.
The details:
OCC Bulletin 2026-54, "Super Intelligence: Supervisory Framework", issued 4 November. "Super Intelligence" is the federal term for AI under Executive Order 14434.
Four illustrative tiers, Critical to Low. "A fully autonomous, customer-facing payment agent sits at Tier 1."
Reviewers may need the "authority to evaluate, reject, modify, pause, or halt" an output. The framework also flags that an agent "could remain available while taking inappropriate or harmful actions."
Non-binding: it "does not establish supervisory expectations for banks", and SI is not a separate risk category. Applies to national banks, federal savings associations and federal branches of foreign banks.
Why it matters: Examiners will now ask which tier each use case sits in and who can stop it. Banks without a consequence-based inventory will be building one under exam pressure. Likely owner: the CRO for the framework, with the head of each business line deploying agents accountable for its tiering.
Across regimes: Supplements existing US model risk guidance; the closest equivalents are PRA SS1/23 in the UK and OSFI E-23 in Canada, both of which cover AI models.
Do now: Map your top ten AI use cases to the OCC's four tiers and name who can halt each one.
🇭🇰 Hong Kong | Banking, payments | Accountability 🟡
Hong Kong wants one named owner for every AI agent
The gap: The HKMA has proposed that every agentic AI deployment has a single accountable owner, and that the bank's accountability cannot be passed to vendors or customers.
The details:
Discussion Paper on Agentic A.I. Governance in the Banking Sector, published 2 November alongside Project SAGE ("Securing Agentic Finance with Governance and Enablement") and a new Agentic Finance Expert Group.
Each use case "should have one designated accountable owner", normally the business owner, with separate technical and control owners.
Agents should act "only in pursuit of an approved business intent, within a defined delegation of authority", with a tested stop mechanism and a possible central registry of material agents.
Non-binding. Responses due 1 February 2027. Scope covers agents banks deploy and external agents acting for customers.
Why it matters: This is the clearest supervisory statement yet that accountability for an agent sits with the business, not IT. Likely owner: the senior executive for the business line deploying the agent, with the CRO owning the control framework.
Across regimes: Close in spirit to the UK SM&CR's named-individual model and to MAS's AI risk management guidelines finalised in October.
Do now: Pick one live agent and write down its business owner, technical owner and control owner. If any line is blank, that is your finding.
🇮🇳 India | Fintech, lending, payments, crypto | Security 🔴
India issues a critical alert on AI-driven attacks against fintechs and lenders
The gap: India's national cyber agency says an AI-assisted threat actor is running attack chains at machine speed against non-bank financial firms, and is returning where clean-up was incomplete.
The details:
CERT-In Advisory CIAD-2026-0058, issued 4 November, severity Critical. It updates CIAD-2026-0046 of 9 October.
Targets NBFCs, digital lenders, fintechs, payment aggregators, PSPs and crypto exchanges. The actor, "CRAFTY FALCON", was named by CSIRT-Fin.
Named tools include the ARTEX family, PentAGI and CyberStrikeAI.
Actions include "Monitor and govern LLM API usage and keys", "Isolate UAT/staging/testing environments from production" and "After any incident, assume persistence".
Why it matters: The advice now includes governing your own AI tools, not just defending against the attacker's. Likely owner: the CISO, reporting to the board's IT strategy committee.
Across regimes: Comparable to the ECB's AI-cyber action plan request to significant institutions and to DORA's ICT risk management requirements.
Do now: Confirm every LLM API key in production has an owner, a scope and monitoring. Revoke the ones that do not.
🇰🇷 South Korea | Banking, insurance, securities, crypto | Security, accountability 🟡
Korea's breach count more than doubles, and CEOs are in scope
The gap: Korea's AI-era hacking wave has reached 37 financial firms, and the regulator is now examining management roles, not just patches.
The details:
FSC release of 5 November, framed around "growing hacking threats driven by artificial intelligence".
40 incidents at 37 firms between 30 September and 3 November; about 473,000 personal records, mostly employee data. Last issue's count was 16 firms.
About KRW 1bn lost from smaller firms' own accounts; no individual users lost money.
Checks on about 580 entities: external access points by end-November, internal controls by year end. Authorities will scrutinise CEO, CISO and CIO roles.
Why it matters: When a supervisor reviews named roles after an AI-assisted incident, the evidence trail becomes the defence. Likely owner: the CEO, with the CISO and CIO for the remediation evidence.
Across regimes: Mirrors DORA, where the management body holds ultimate responsibility for ICT risk, and the UK SM&CR's named senior manager model.
Do now: Check your CISO and CIO reporting lines are documented and that the board has seen your external access point inventory this quarter.
The holiday streaming season is almost here
Are you ready? As revealed in our holiday playbook, Roku Ads Manager lets you easily set up, measure and optimize your self-serve holiday ad campaign. You can even create a shoppable ad unit that lets streamers buy directly from their remote.
🇬🇧 UK Desk
AI, risk and accountability across UK financial services this week.
FCA finds untested kill switches and incomplete AI inventories at wholesale banks
The FCA's Autumn 2026 wholesale banks supervision insights, published 2 November, report internal audit findings including "incomplete AI inventories", conditional approvals recorded as full approvals, and "no documented list of authorised individuals" to operate kill switches. In one incident, "inaccurate AI-generated code recommendations were not identified during the human review stage", causing a loss of service. FCA
What it means for accountable firms: the FCA is reading internal audit's AI findings, so who can stop each model needs to be written down, not assumed.
UK's five largest banks reported to restrict AI agents' account access
The Telegraph reported on 8 November that HSBC, Barclays, Lloyds, NatWest and Nationwide are blocking or restricting AI agents from accessing customer accounts. It follows similar moves by ten North American banks covered in our last issue. Telegraph
What it means for accountable firms: a block is a policy decision, so it needs a named owner, a documented rationale and a way to tell agent sessions from customers.
Lender lens: FCA sets terms for motor finance settlement offers
There was no AI-specific lender news this week. On 3 November the FCA said firms may make settlement offers while legal challenges to the redress scheme continue, "as long as customers are offered no less than they would receive through the scheme", and that these offers cannot be referred to the Financial Ombudsman Service. FCA
What it means for lenders: if offers are calculated at scale, each one must be provably no less than the scheme amount.
Coming up
🛠️ Practitioner's corner: tier your AI agents by consequence in one afternoon
List what can act. Every AI system that can send, pay, approve, release or contact a customer without a human pressing the button.
Ask the failure question. If it did the wrong thing 1,000 times before anyone noticed, what breaks? Money, customers, markets or nothing external?
Assign a tier. Use four bands, Critical to Low. Irreversible and customer-facing goes to the top, regardless of test results.
Name three people. Business owner (accountable), technical owner, control owner. The control owner must be independent of the build.
Test the stop. For each Critical agent, have the named person pause it in a test environment and record how long it took.
⚡ Quick hits
UK & Europe
🇬🇧 [Cross-sector] The ICO's agentic AI call for evidence, covering controller accountability and automated decisions, closes 20 November. ICO
🇨🇭 [Banking, fintech] FINMA's revised Circular 2016/7 on video and online identification took effect 1 November, adding liveness detection with AI risks cited. FINMA
🇩🇪 [Banking] Deutsche Bank says building governance in from ideation cut idea-to-production time "from around 12 months to as little as three months". Deutsche Bank
Americas
🇺🇸 [Fraud] SDNY charged three men over a "more than $129 million" scheme using AI to impersonate executives and lawyers; victims include a UK fintech ($11.6m), a Massachusetts mutual fund company and a Rhode Island bank. DOJ
🇺🇸 [Banking] Comments on the interagency third-party risk management proposal (OCC, Fed, FDIC, NCUA) are due 16 November. Fed
Asia-Pacific & Middle East
🧰 RegTech & security watch
Google and Mandiant report "LLMJacking", with stolen cloud credentials used to run and resell firms' AI access; financial services were among affected sectors, with charges reaching $1 million at some organisations. Control risk: an unmonitored model endpoint is now a fraud and cost exposure, not just a data one. Google Cloud
Revolut has reportedly moved coding work from a commercial model to self-hosted open-weight models, with reported inference savings of up to 90%. Control risk: self-hosting moves safeguards, patching and abuse monitoring onto the firm. TechRadar
Mistral added managed agents to its Studio platform, naming BNP Paribas and ABANCA among production users. Control risk: a managed agent is still the bank's agent; vendor change notices need a named recipient. Mistral
📅 Calendar
16 November | US interagency third-party risk management guidance: comments due
20 November | UK ICO agentic AI call for evidence closes
24 November | Wells Fargo online access agreement with AI-tool terms takes effect
30 November | Korea FSC external access point checks on about 580 financial entities due
📊 Pulse check
Who is the accountable owner for your highest-risk AI agent today?
RegTech Spotlight - BuildMoat
Disclosure: BuildMoat publishes The Accountability Record. This is our own product slot, kept separate from editorial. Evaluate it as you would any vendor.
The OCC wants to know who can halt an agent. The HKMA wants one named owner for each. BuildMoat works on that gap: named accountable approval before AI operates, control while it runs, and an evidence ledger of what happened. One of the components that allow us do this is the Authority Studio.
Authority Studio is where an AI agent gets its permission to act, before it goes live.
Before deployment, the firm defines exactly what each agent is allowed to do. That covers its mandate, its financial limits, the data it can touch, the systems and endpoints it can call, and the points where a human has to review its decisions. A named accountable senior manager then signs off that set of rules.
Everything else in BuildMoat checks against this. While the agent runs, every action is compared with what was authorised. Anything outside the limits is blocked, escalated or flagged, and the record of who approved what sits in the Evidence Ledger.
In short, if a regulator asks “who allowed this AI to do that, and on what terms?”, Authority Studio is where the answer was set, ahead of time and with a name attached.
Learn more: buildmoat.co.uk
What did you think of today's edition?
That’s it for the week!
If something landed, reply and tell me. If something was wrong, reply and tell me that too. I read every response.Forward this to one person who should be reading it.
For sponsorship enquiries, email [email protected]
Until next week.
Emeka Ogbonnaya
Founder BuildMoat