data sovereignty

Bay Street's Vishing Breach Exposes Shared-Cloud Legal AI Risk

Four Bay Street firms lost 31,000+ files to a vishing attack on cloud identity, not an AI tool. Here's why architecture, not AI, decides blast radius.

RAGbase Legal Research TeamNovember 4, 2026 10 min read

On October 19, 2026, an employee at McCarthy Tétrault took what looked like a routine IT help-desk call. By October 23, four Bay Street firms — McCarthy Tétrault, Bennett Jones, BLG, and Davies — were confirming breaches. The total damage: at least seven compromised employee accounts, client revenue reports, mandate descriptions, and more than 31,000 client files exfiltrated from a single firm. No malware. No zero-day exploit. No AI model was hacked. The attackers simply called, impersonated IT support, and convinced someone to approve a login.

This is the story the legal industry needs to sit with before its next AI procurement cycle, because the instinct after a breach like this is to ask "which software failed?" The more useful question is "which architecture made failure catastrophic instead of contained?" At Bennett Jones, the firm's core document management system was never touched — the compromise hit peripheral cloud and identity layers. That single detail is the entire lesson. The attack surface that mattered wasn't a specific application. It was the centralized identity perimeter sitting underneath dozens of SaaS tools, including, increasingly, AI copilots.

What Actually Happened: Anatomy of a Vishing Campaign

According to reporting from The Globe and Mail, the attackers ran a coordinated five-day campaign (October 19-23, 2026) targeting employees across four of Canada's largest law firms simultaneously. The mechanics were consistent across targets:

  • Pretext: Callers impersonated internal IT help-desk staff, a pretext that works precisely because it mirrors legitimate password-reset and MFA-reenrollment workflows employees are trained to comply with.
  • Real-time MFA relay: Rather than trying to defeat MFA outright, attackers captured credentials and one-time codes as victims entered them on rotating lure domains, then relayed them to the real login portal within the live authentication window — effectively hijacking the session token, not just the password.
  • Rotating lure domains: Each attempt used a fresh domain, limiting the effectiveness of static blocklists and giving defenders no time to flag and propagate indicators before the next target was hit.
  • Scale per firm varied, severity did not: McCarthy Tétrault lost four accounts and 31,000+ files; Bennett Jones lost two accounts' worth of client revenue reports and mandate descriptions; Davies lost one account. The number of compromised credentials was small. The downstream exposure was not.

That asymmetry — a handful of phone calls producing tens of thousands of exposed files — is the structural story, not the social-engineering story. Vishing is not new; it is the same technique behind the 2023 MGM Resorts breach (estimated at over $100 million in losses) and the 2022 Uber intrusion via MFA fatigue and help-desk impersonation. What changes the outcome each time isn't the sophistication of the call. It's what the compromised credential unlocks.

The Real Lesson: It's the Identity Layer, Not the Application

Every firm in this breach almost certainly has layered security: endpoint detection, DLP, SIEM, security awareness training. None of it stopped a well-executed phone call because the call didn't target a technical control — it targeted the single identity that, once authenticated, had standing access to firm-wide cloud resources. This is the defining feature of modern SaaS architecture: one set of credentials, federated across dozens of connected tools, each inheriting the same trust.

That is also, structurally, how most AI copilots are deployed today. A firm connects an AI assistant to its document management system, email, and collaboration tools via OAuth and single sign-on. The AI tool itself may be well-engineered and well-secured — but it sits inside the same identity perimeter as everything else. If that perimeter is breached via vishing, the AI layer isn't a separate risk category; it's one more door that opens with the same stolen key.

Bennett Jones' experience is the clearest proof point here. Its document management system — the actual repository of privileged client work product — was not breached. The compromise hit the cloud identity and peripheral SaaS layer around it: account access sufficient to pull revenue reports and mandate metadata, but not sufficient to reach the core repository. That's not luck. That's a difference in how tightly coupled the identity layer was to the most sensitive systems. Firms that federate AI tools, file sync, and document repositories under one broad identity pool are, by design, making that coupling tighter — and making a single vished account more valuable to attackers.

Centralized SaaS vs. Private AI: Comparing Blast Radius

The question for a managing partner or CIO isn't "could my firm be vished?" — assume yes. The question is: if one employee's session is hijacked tomorrow, what is the maximum exposure?

DimensionShared-Cloud Legal AI (typical SaaS deployment)Private/On-Prem AI (private AI deployment)
Identity perimeterFederated SSO spanning email, DMS, AI tool, and other SaaS apps under one token-issuing authorityFirm-controlled identity system; AI access scoped separately from vendor-hosted session pools
Blast radius of one hijacked sessionPotentially firm-wide, across every connected app the account can reachLimited to the firm's own infrastructure and access-control layer, not a shared vendor tenant
Session tokens at riskVendor-side session tokens, reusable across the vendor's multi-tenant cloudNo external cloud vendor session to hijack for the core retrieval/document layer
Document corpus locationIndexed and often cached within vendor's cloud environmentFull corpus, vector store, and retrieval index remain on firm-controlled infrastructure
What leaves the building per queryEntire documents or large context windows, depending on vendor architectureOnly the minimal retrieved chunks needed to answer a given question, sent to the selected LLM under the firm's API terms
Access logsHeld by vendor, subject to vendor's retention and disclosure policiesHeld on firm infrastructure, under firm's own audit and retention controls
Attack surface added by AI adoptionNew connectors, new OAuth scopes, new centralized identity dependencyAI usage logs and document access confined to infrastructure already governed by firm security policy

The point isn't that cloud SaaS AI tools are reckless. Harvey, CoCounsel, Lexis+ Protege, Legora, and Claude Cowork are built by serious engineering teams with real security programs. The point is architectural: every one of them, by design, requires the firm's documents and identity to interact with a multi-tenant cloud environment that is now part of the firm's attack surface. A vishing campaign that compromises firm-side credentials doesn't need to breach the AI vendor at all — it just needs the AI tool to already be connected, authenticated, and trusted.

What Changes With a Private AI Architecture

The honest comparison here is not "RAGbase never sends data out, the others do." RAGbase Legal also calls LLM providers — Claude, GPT-class models, or others, depending on the firm's chosen configuration. The meaningful difference is where the heavy infrastructure lives.

In a RAGbase deployment, the agentic scaffolding — connectors, retrieval layer, vector store, permissioning, logs, and workflow orchestration — runs on the firm's own infrastructure, not a vendor's shared cloud tenant. When an associate runs a query through case search or a research workflow, the system retrieves the relevant passages from the firm's own indexed corpus, and only those minimized chunks — not the full documents, not the full matter file — are sent to the selected model provider to generate a response. The 31,000 files that left McCarthy Tétrault's environment were whole files, pulled through a compromised account with broad standing access. In a retrieval-scoped architecture, there is no equivalent single point that yields a comparable volume, because the AI layer itself never holds a centralized, broadly-accessible copy of the firm's entire document universe — the firm does, under its own identity controls.

This also changes what a vished credential is worth to an attacker. If help-desk impersonation compromises an employee account in a firm running AI on its own infrastructure, the attacker still has to deal with the firm's native access controls, segmentation, and on-site logging to get at anything resembling 31,000 files — the same controls that would apply whether or not AI was ever introduced. There's no additional vendor-hosted session pool layered on top that broadens what a single stolen login can reach.

What CIOs Should Actually Ask Vendors Now

This breach should reshape AI procurement conversations at AmLaw 200 firms, not toward paranoia about AI specifically, but toward precise questions about identity architecture:

  • Where does the retrieval index live? If it's in the vendor's multi-tenant cloud, a compromised firm credential plus a vendor-side vulnerability (or a compromised vendor employee account) both become relevant risks.
  • What's the minimum unit of data exposure per query? Full-document retrieval and full-context uploads create larger exposure per compromised session than scoped, chunk-level retrieval.
  • Are AI tool sessions part of the same SSO/token pool as email and file sync? If yes, a vished help-desk call targeting any one of those systems is effectively targeting all of them.
  • Who holds the access logs, and for how long? Firm-controlled logging matters for incident response speed — Bennett Jones and McCarthy Tétrault needed to reconstruct exactly what each compromised account touched, fast.
  • Can the AI deployment be segmented from the broader identity perimeter? Firms should be able to answer this before signing, not after an incident.

Our AI for law firms guide walks through this evaluation in more depth, including where retrieval-scoped architectures fit relative to full-cloud SaaS platforms for different risk tiers of work.

Forward-Looking: Vishing Will Get Worse Before Governance Catches Up

Real-time MFA relay and rotating lure domains are not a one-off tactic — they're a maturing playbook, and AI itself is making vishing more effective, not less. Voice cloning and LLM-generated pretexting scripts lower the skill floor for attackers running exactly this kind of multi-target campaign. Expect more coordinated hits across multiple firms in compressed windows, because the economics favor it: the same help-desk pretext and relay infrastructure can be reused across targets with minor variations, as this campaign demonstrated across four firms in five days.

The firms that fare best in the next wave won't necessarily be the ones with the newest AI tool. They'll be the ones that can answer, with precision, what a single compromised credential can actually reach — and have deliberately shrunk that answer. For sovereignty-critical workloads — privileged matter files, M&A due diligence rooms, regulatory investigations — that argues for keeping the retrieval layer and document corpus inside firm-controlled infrastructure, using cloud-hosted AI tools where the risk profile allows it, and treating the two as complementary rather than an all-or-nothing choice.


If your firm is evaluating how AI connects to your document systems, the question worth raising in the next security committee meeting isn't which AI vendor has the best features — it's which architecture limits what a single vished help-desk call can expose. That's a conversation about identity design as much as it is about AI, and it's worth having before the next coordinated campaign, not after.

Frequently Asked Questions

Was the Bay Street law firm breach caused by an AI tool?
No. The attackers used vishing (voice phishing) and real-time MFA relay to compromise individual employee accounts tied to firm-wide cloud identity systems between October 19-23, 2026, affecting McCarthy Tétrault, Bennett Jones, BLG, and Davies. No AI platform was involved in the initial compromise — the entry point was centralized cloud identity infrastructure, not an AI vendor or model.
Why didn't multi-factor authentication (MFA) stop the attack?
The campaign used real-time relay techniques and rotating lure domains, meaning attackers captured and immediately replayed valid MFA codes or session tokens as employees entered them on spoofed help-desk pages. This bypasses conventional push/SMS MFA because the authentication itself is legitimate — it's the session token, not the credential, that gets hijacked.
Does on-premise or private AI eliminate this kind of risk entirely?
No single architecture eliminates social engineering risk, since employees can still be targeted directly. But on-prem/private AI reduces blast radius by keeping the document corpus, retrieval index, permissions, and access logs on firm-controlled infrastructure rather than a shared cloud identity pool spanning thousands of client files, so a single compromised account cannot cascade into tens of thousands of exposed files the way it did at McCarthy Tétrault.

Related Articles

R
RAGbase Legal Research Team
Research

RAGbase builds private AI systems for law firms: deployed on the firm's own infrastructure, zero data retention, full ownership.

See How RAGbase Works on Your Data

30-minute call. We scope your use case and show the system live.

We use audience and marketing cookies (Google Analytics, LinkedIn). No tracker loads without your consent. Learn more