Artificial intelligence can help a team examine alerts. But when it accesses an organisation's documents and applications, it also becomes part of its information system. Security now concerns both the assistant and the data it handles.

A recent report with a defined scope

In its German presentation posted on 2 October 2026, with a statement dated 1 October, Microsoft describes growing attack automation. The report mainly covers July 2025 to June 2026, unless otherwise stated.

These observations come from a security technology provider. They shed light on the threats it observes, but are not a comprehensive census of all attacks affecting all businesses.

Server racks in a computer room
Illustrative photo: Jemimus, 2008 — CC BY 2.0; orientation corrected.

Useful tools can create exposure

Microsoft also identifies threats targeting AI models and applications. The problem is therefore no longer limited to a more persuasive fraudulent message: a connected tool can process untrusted information while retaining access to internal resources.

Our analysis: the decisive question becomes how much power the agent actually has. The quality of its answers is insufficient to assess the risk arising from its permissions and actions.

Three separate challenges for organisations

NIST's draft Cyber AI Profile distinguishes three areas: securing AI components, using AI for defence and confronting AI-enabled attacks. The page consulted presents an initial draft published on 16 December 2025, not a security certification.

This distinction helps explain why purchasing a cybersecurity assistant does not, by itself, protect every other AI use within the organisation.

Business continuity remains the intended outcome

Our analysis: faster alert analysis only has value if it improves decisions and reduces the consequences of incidents. Visibility into agents, control over access and people's ability to take back control will form part of that assessment.

Conclusion: measure control as well as speed

The report encourages treating AI as both a defensive tool and an asset that needs protection. For businesses, progress will be measured by the resilience achieved, rather than the number of actions automatically assigned to software.