← Back to feed

The same class of AI that regulators warn could arm attackers is being deployed at scale across the financial services sector. The gap between those two positions is narrowing fast.

Australia’s Prudential Regulation Authority issued an advisory to the banking sector warning that frontier AI models, specifically citing Claude Mythos-class capability, could provide attackers with advanced technical capabilities that most financial institutions are currently unequipped to defend against. APRA’s framing was direct: the threat is not hypothetical, and banks should be accelerating internal capability rather than waiting for regulatory guidance.

The advisory lands as the largest cloud-based accounting software platform in Australia and New Zealand moves in the opposite direction.

Xero Signs Multi-Year Claude Deal

Xero, which serves more than four million small business customers globally, announced a multi-year partnership with Anthropic to integrate Claude across its accounting platform. The stated goal: transform accounting from a manual process into an automated, agentic workflow.

Under the partnership, Claude will handle reconciliation prompts, compliance drafting, supplier payment categorization, and workflow routing for Xero’s SMB customer base. Anthropic described the deal as one of the largest agentic deployments in the accounting software category.

Xero has not disclosed the financial terms of the agreement or specified which Claude model will be deployed initially, though the partnership framing suggests access to models beyond the current public API tier.

The Regulator’s Concern

APRA’s warning is narrowly focused on offensive capability, not enterprise deployment risk. The regulator’s concern is that Claude Mythos-class models, operating at or near frontier capability, can assist attackers in writing novel malware, analyzing financial system APIs for exploitable conditions, and generating highly convincing social engineering materials at a scale that exceeds current fraud detection thresholds.

APRA said the banking sector has had meaningful exposure to AI-assisted cyber threats for at least 18 months, but that the capability step represented by the current frontier models is qualitatively different. The advisory recommended that banks request expedited access to frontier AI tools for internal red-teaming and defensive capability building.

That recommendation puts APRA in the same position as several other national financial regulators: the best defense against AI-enabled attacks is familiarity with the same tools attackers will use.

The Enterprise Adoption Gap

The Xero-Anthropic deal illustrates the deployment velocity that makes APRA’s advisory difficult to operationalize. Xero serves businesses that are largely not regulated by APRA and have no requirement to conduct AI security assessments before deploying agentic tools. Four million small businesses handling payroll, tax, and payment workflows through a Claude-powered interface is a significant expansion of the frontier model attack surface.

The accounting software category is not the only vector. In the same week, Claude was integrated into Microsoft Word for legal document drafting, Autodesk Fusion for design workflows, and several enterprise financial platforms through direct API deals. The enterprise deployment pace substantially exceeds the rate at which financial sector security teams can build defensive understanding of what frontier models can do.

APRA did not issue binding requirements alongside its advisory. Banks that take no action face no immediate regulatory consequence. The advisory is a signal, not a mandate.

How that changes, and how fast, is now the open question for the sector.