When AI systems are compromised, detection without recovery is not resilience

Can businesses determine what happened, contain the consequences and safely reverse unintended outcomes?

Summarise
阅读简体中文版 (beta)
    • As AI moves from assisting employees to acting on their behalf, the nature of operational risk changes.
    • As AI moves from assisting employees to acting on their behalf, the nature of operational risk changes. IMAGE: REUTERS
    Published Thu, Nov 19, 2026 · 03:00 PM

    RECENT data breaches involving MoneySmart, M1 and Doctor Anywhere have put cyber risks back in focus for Singapore businesses.

    On Nov 13, the Cyber Security Agency of Singapore urged all organisations to “take immediate steps to strengthen their cyberdefences”.

    While most conversations have focused on how these incidents happened and what information was exposed, they raise a bigger, more urgent question for business leaders. If critical systems are hit, how long can the business afford to be out of action before the financial, operational and reputational damage begin to pile up?

    Consider two incidents with very different outcomes.

    In September 2026, Boustead Singapore disclosed that an overseas business unit had isolated affected systems, activated business continuity procedures and recovered critical data and systems from backups following a cyberincident. Its operations suffered no material disruption.

    Contrast that with the 2025 cyberattack on Marks & Spencer, which forced systems offline and disrupted online operations for weeks.

    While physical stores continued trading, customers could not access certain online shopping services, with operations returning in phases.

    The circumstances were different, but both incidents highlight a key point. Getting systems back online is one thing; knowing whether the business can safely resume operations is another.

    As organisations give artificial-intelligence agents access to sensitive systems, leaders must ask: Can the team determine what happened, contain the consequences and safely reverse unintended outcomes?

    Commvault’s 2026 State of Data Resilience report found that only 45 per cent of organisations in Singapore are very confident that they could identify a compromised AI system.

    If organisations are not even confident they can identify a compromised AI system, determining what has been affected, what needs to be recovered, and whether the team can restore it to a clean, trusted state could be even more challenging.

    AI changes the nature of the blast radius

    As AI moves from assisting employees to acting on their behalf, the nature of operational risk changes.

    An autonomous AI agent may make decisions and take actions across multiple connected systems using identities and permissions legitimately granted by the organisation.

    This autonomy creates business value, but it also means that a compromised identity, inaccurate data, incomplete instruction or poorly designed workflow can trigger consequences across data, applications and business processes before a person has time to intervene.

    The challenge then is to understand the chain of actions that followed.

    What data did the system access? What decisions did it influence? Which transactions did it initiate? What downstream processes accepted its output? Which actions need to be reversed, reconciled or reviewed. Without that understanding, an organisation cannot know whether the incident is truly contained.

    And if it cannot establish which data and systems remain trustworthy, restoring the affected technology may not be enough.

    Recover what the business needs, then prove that it works

    No organisation can restore everything at once. Recovery, therefore, needs to begin with understanding which capabilities the business needs first, and what those capabilities depend on.

    The minimum viable company, or MVC, defines the systems, data and operations that must remain available, or be restored first, for the organisation to continue functioning during disruption.

    An MVC differs by organisation. For a bank, it might include enabling customers to make time-critical payments and withdraw funds.

    For a logistics provider, it might mean accepting shipping instructions, tracking goods in transit and releasing deliveries at their destination.

    The priority is to determine which business processes must continue or be restored first, and then map the systems, data, people and other dependencies needed to support them.

    The same principle applies to AI. One business process supported by AI may require rapid recovery, while another one may tolerate a longer interruption.

    Prioritisation should therefore be based on business impact and recovery requirements, not simply on whether a workload uses AI.

    It must also account for the identities, applications, infrastructure and data required to restore the complete business process safely.

    Boards do not need to know every recovery dependency or direct the operational response.

    But they should require evidence that management has identified the organisation’s critical business capabilities, and agreed on the order in which they will be restored cleanly and tested so that they can operate safely following disruption.

    The next test for trusted AI: Recovery readiness

    As Singapore advances trusted AI, resilience must become part of that trust. Guardrails matter, but so does knowing the business can recover when those are not enough.

    As AI moves from assisting to acting, recovery readiness must become part of AI readiness.

    Singapore businesses must be able to demonstrate that they can contain unintended consequences, restore critical operations and recover safely when existing safeguards fall short.

    The writer is field chief technology officer for Asia-Pacific at Commvault

    AI tools were used to support research, framework development and proofreading of the article before submission