🇸🇬Singapore·AI News·12 Nov 2026·via Fintech News Singapore

    Can Banks Safely Restart Their AI Agents After a Cyberattack?

    French lender Crédit Mutuel Arkéa faced a cyberattack resulting in an estimated €100 million missing, Reuters reported. The incident highlights complex challenges for banks, particularly how to safely restore AI agents used for fraud detection. Experts warn that traditional recovery methods might miss critical details, potentially leaving systems vulnerable even after a restore.

    Restoring AI agents after a cyberattack presents a challenge beyond simply recovering a model. Commvault's June 2026 report, "The AI Resilience Blueprint," identifies this as a "context gap." This occurs when a model backup might be an older version, while the script for preparing its data was updated later. Recovery teams then must determine if these components ever operated together. Furthermore, critical elements like account notes may reside in separate databases or self-hosted model registries, potentially outside the scope of a model's backup coverage.

    A significant risk lies in how false details can persist or propagate within these complex systems. An altered note, for instance, can pass through multiple agents, with one pulling transaction history and another writing a summary. Agents remembering earlier tasks may carry such details into later cases, and with write access, they could save tampered information into other systems. Akto’s 2025 State of Agentic AI Security survey found that only 17% of surveyed organizations continuously monitored exchanges between agents, indicating limited visibility into these handoffs. Moreover, restoring an AI model and its data might not revert an attacker's changes to the agent's permissions, allowing it to retain unauthorized access.

    To address these issues, the Commvault report suggests banks trace the source and handling of training data, as tampering can alter model behavior. It also proposes using a system of record for AI agent recovery to tie together model versions, data, instructions, and permissions at a specific point in time. This approach, paired with logs of agent activity, could help identify when an altered note first appeared. Deloitte’s 2026 research indicates that only about one in five of 3,235 business and IT leaders surveyed felt their organizations had a mature approach to governing autonomous agents, underscoring the broader challenge.

    Before an AI agent is reconnected to customer systems, it is crucial to verify that all components form a "known-good state." This involves checking restored permissions against originally approved access and testing the agent's performance with known fraud cases. Such recovery drills, as recommended by the Commvault report, can help confirm the integrity of the restored setup and provide a verifiable basis for putting agents back to work, particularly when facing scrutiny from boards, auditors, or regulators.

    Share this article

    Original reporting by Fintech News SingaporeWe don't republish, read the full story →

    Related reading

    6 stories