As of November 2026, attackers use AI to impersonate executives, steal credentials at scale and target the AI tools firms rely on. Super Chain’s view: a private LLM keeps client data, keys and the audit trail on infrastructure the firm controls, which shrinks what can leak. It is not immune; deepfake calls and compromised packages need their own controls.
Mostly to impersonate people and to work at a speed no human team matches. Cloned voices and deepfakes copy the executives who approve payments, and AI agents work through stolen credentials in hours rather than weeks.
A private bank in Europe lost a reported €95 million after fraudsters used an AI-cloned voice of a senior executive (reported by Reuters via Biometric Update, October 2026).
In the UK, Report Fraud and the City of London Police counted £24.3 million stolen between January and August 2026 through investment schemes that used AI, including deepfakes, across more than 1,300 reports (City of London Police, 29 October 2026).
Anthropic reported in September 2026 that one attacker’s AI agents dumped more than 2,100 Azure AD token sets across over 40 corporate tenants in about 34 hours (reported by Unite.AI, September 2026).
Because they sit where client data and credentials meet. An AI assistant reads documents, holds API keys and often has access to email or files, so compromising the tool can expose everything it touches.
On 21 October 2026, a compromised release of LiteLLM, an LLM gateway with about 23 million weekly downloads, shipped malware that steals cloud keys and AI-tool credentials (Help Net Security, 22 October 2026).
In 2025, EchoLeak (CVE-2025-32711) let an attacker use AI command injection in Microsoft 365 Copilot to disclose information, and Wiz Research found a publicly accessible DeepSeek database exposing over a million lines of logs, including chat history and secret keys (Wiz Research, January 2025).
Unapproved tools add to it. IBM reported that one in five organisations had a breach due to shadow AI, and that high shadow-AI use added an average of $670,000 to breach costs (IBM, 30 July 2025).
In the UK, financial services was the costliest industry for a breach in 2025, at an average of £5.74 million (IBM UK, 30 July 2025).
It changes custody. With a private LLM, the documents, prompts, model and audit log run on infrastructure the firm controls, isolated from other organisations, so client data is not sitting in a vendor’s tool the firm cannot inspect.
It does not make the firm immune. Three limits matter:
All three markets now name AI-enabled attacks or AI data leakage directly, and expect firms to show controls rather than intentions. The rules differ in form, but each asks who controls client data and how fast a firm can detect and report a breach.
Hong Kong. The HKMA’s 3 September 2026 circular on remote on-boarding warns that AI-driven automation and deepfakes have increased the potential for impersonation and online fraud at much greater scale (HKMA, September 2026).
In Hong Kong, the SFC’s circular on generative AI names heightened risks of cyberattacks and inadvertent leakage of confidential information (SFC, 12 November 2024), and the PCPD’s AI Model Personal Data Protection Framework (June 2024) sets out how to procure and use AI with personal data.
United Kingdom. A joint statement from the FCA, the Bank of England and HM Treasury says firms that have underinvested in core cyber security fundamentals are likely to become progressively more exposed as frontier AI cyber capabilities grow (FCA, 15 May 2026).
United States. The SEC’s FY2027 examination priorities add a focus on AI-driven cyber threats, reported by National Law Review on the SEC release of 23 October 2026.
In the US, amended Regulation S-P requires customer notice no later than 30 days after a firm becomes aware of an incident (SEC, 16 May 2024).
For the deployment decision by market, read on-premise AI for wealth management.
Seven steps, in order. The first three decide most of the exposure before any software is bought; the rest keep it closed once AI is in use.
Super Chain is a private AI platform that runs on the firm’s own servers or private cloud, isolated per organisation, with answers cited to the firm’s own documents. It covers the custody, permissions and audit steps above; the payment and staff controls stay with the firm. The trust and security page sets out the architecture, and five controls for using AI safely and the cyber-safe AI roadmap set out the order of decisions.
This guide is for asset and wealth managers, family offices, brokers and banks in Hong Kong, the UK and the US whose staff already use AI or are about to, and who hold client data that must not leak.
It is not a full cyber programme. Email security, endpoint protection and penetration testing sit outside it, and a firm needs those whether or not it uses AI.
Super Chain is accountable for the platform and its configuration. The office remains accountable for which documents go in, who may see them, and the terms of the model provider it chooses.
Super Chain is built by Super Cat Technology. The same team built a fully on-premise retrieval system for a Big 4 firm, where every answer returns with its source paragraph.
Partly. A private LLM keeps client documents, prompts, keys and the audit trail on infrastructure the firm controls, so less data sits in tools the firm cannot see. It does not stop a deepfake call or a compromised software package. Those still need process controls, such as call-back checks, and dependency vetting.
Through impersonation and scale. In 2026, reported cases include an AI-cloned executive voice used to move €95 million out of a private bank, and UK police counted £24.3 million lost to AI-enabled investment fraud between January and August. Attackers also aim AI agents at stolen credentials.
Because they hold client data and credentials. In October 2026 a release of LiteLLM, a widely used LLM gateway, shipped malware that steals cloud and AI-tool credentials. Earlier, a Microsoft 365 Copilot flaw (CVE-2025-32711) allowed information disclosure, and an exposed DeepSeek database leaked chat history and secret keys.
Each names AI and cyber risk directly. The HKMA warned in September 2026 that AI and deepfakes raise impersonation risk at scale, and the SFC flags data leakage from generative AI. UK regulators warn under-invested firms will become more exposed. The SEC added AI-driven cyber threats to its FY2027 exam priorities.
With an inventory of where staff already paste client documents into AI tools. That shows the exposure before any purchase. Then decide which data may reach a hosted tool, move the rest to AI on infrastructure the firm controls, pin AI dependencies, and add call-back checks for payment instructions.
We load one vehicle’s documents onto a dedicated, isolated deployment hosted by Super Chain, with an enterprise model API. Your team asks real questions of them for 30 days, and each answer comes back cited to the page it came from. Production deployments run on-premise, on the office’s own infrastructure.
Choose a time below to scope the pilot with us. If you would rather see the platform first, book a private demonstration.