Evidence note: More than 100 companies warned in August 2026 that AI-enabled cyberattacks could become more widespread. This was an industry forecast and call to action, not evidence that every predicted attack has occurred.

AI gives organisations leverage. Attackers can gain leverage too. Security has to sit inside the AI operating model from the first architecture decision.

The warning is about preparation

Reuters reported that companies including major AI labs, cloud providers, security firms and financial organisations called for a society-wide defensive surge. Their position is that increasingly capable models can increase the speed and scale of offensive activity.

Five controls that cannot wait

Identity and permissions

Know which person, service or agent initiated every material action. Use least privilege, isolated credentials and rapid revocation.

Architecture and segmentation

Limit the blast radius between tools, data, customers and environments. A convenience connection should not become an invisible trust relationship.

Monitoring and immutable logs

Detect unusual access and preserve evidence outside the control of the process being monitored.

Deployment control

Test capability changes in contained environments with an approved route to release, hold and rollback.

Human incident authority

Name the people who can contain the system, communicate the incident and make the recovery decision.

Governance is part of defence

A security tool cannot compensate for unclear ownership. SOS treats permissions, evidence, change control and human authority as buyer questions, not technical footnotes.

Sources reviewed

Want your AI deployment reviewed against the same governance controls?

Talk to us about AI governance