Connect with us

Tech

Microsoft Warns AI Agents Are an Insider-Risk Problem: Four Security Priorities for Businesses

Microsoft argues that AI agents need visibility and strict permissions, while businesses must continue patching vulnerabilities and removing outdated systems.

Published

on

Buildings and landscaped plaza on Microsoft campus in Redmond

TECH,- As companies deploy AI agents inside business systems, cybersecurity increasingly depends on knowing what those agents can access and what actions they can take without supervision. In a 19 November security essay, Microsoft argued that organisations must strengthen identity controls, patching and monitoring rather than rely on AI tools to compensate for basic weaknesses. The warning is relevant to businesses experimenting with autonomous workflows, where a trusted-looking application can cause harm even without a conventional malware infection.

An AI agent may search a document repository, read customer records or use software tools to carry out instructions. Those permissions can save employees time, but they also create a new pathway for mistakes and manipulation. A malicious prompt embedded in an untrusted document, for instance, might try to persuade an assistant to disclose information or act beyond the user’s intended purpose.

Why Microsoft calls for more visibility

Microsoft says company boards should understand how many agents operate inside their organisation, which resources they can reach and which decisions they can make independently. Many companies already struggle to inventory conventional applications and user accounts, so adding agents without registration or monitoring can make the problem worse.

The company describes four areas for immediate action: reducing accumulated technical debt, closing gaps in security patching, gaining broad visibility into systems and using AI to defend at machine speed. These are recommendations rather than a government regulation or proof that Microsoft’s products alone can prevent incidents.

Technical debt includes ageing systems, outdated software and default passwords that organisations have postponed addressing. Such weaknesses remain attractive to attackers because sophisticated automation cannot repair every unsupported component or compensate for excessive privileges.

Why faster patching has become more important

When software developers publish a fix for a vulnerability, defenders must install or deploy it, while attackers may study the same information to understand the weakness. AI can potentially accelerate both sides of that contest. Microsoft therefore recommends measuring the time it takes to fix problems, not simply the number of vulnerabilities recorded in a dashboard.

That recommendation matters operationally. A company can report thousands of identified issues yet leave its most dangerous exposure open for weeks. Prioritisation should combine severity, exploitability, business criticality and actual access paths. Automated remediation can help, but changes to production systems need testing and rollback plans.

Treat agents as privileged identities

An AI assistant connected to payroll or a customer database should not automatically inherit broad administrative authority. Principles such as least privilege, strong authentication and logging apply to software agents as well as employees. High-impact actions may need explicit approval, transaction limits and a clear method for disabling an agent quickly.

The security problem also spans vendors. Enterprises may combine AI models, workflow platforms, data stores and cloud services from several suppliers. That makes it important to understand how authorisation travels between systems and which organisation is responsible when a connected service fails.

What businesses in Asia should do first

In Singapore and across ASEAN, regulated companies must consider local information-security and data-protection obligations. Yet the general actions apply more widely: maintain an accurate inventory, limit agent permissions, check logs, test unusual instructions and establish human review for consequential decisions. Staff need guidance on when an AI system is permitted to use confidential data.

Microsoft points to its own security and agent-management offerings as possible tools, but product adoption should follow an assessment of the organisation’s risks. No purchase can replace good governance or a disciplined response to incidents.

The lesson is not to avoid AI agents altogether. It is to recognise that giving software the ability to act creates responsibilities beyond checking whether its answers sound useful. Businesses should demand measurable controls and maintain the ability to explain, review and stop automated decisions.

Copyright © 2020 Todayinasian.com