Artificial intelligence is no longer merely an experimental security concern. New findings highlighted by IBM indicate that more than one in five surveyed UK organizations experienced an AI-related breach during the last year, while AI-driven attacks increased by 56 percent over the same period.
The average financial impact associated with an AI-related breach was reported at approximately $6 million. Although individual organizations will experience very different costs, the figure reflects a broader reality: poorly governed AI adoption can create access paths into sensitive data, cloud infrastructure, software development systems, and business workflows.
The AI Model Is Not Always the Weakest Link
Attackers frequently targeted the supporting technology around AI rather than directly defeating a model. Compromised APIs and plugins accounted for 27 percent of reported targeting activity, while cloud misconfigurations represented another 27 percent. These are familiar security weaknesses, but AI can amplify their impact because models and agents often connect to large data stores and privileged business applications.
AI is also changing the attacker's side of the equation. Deepfake impersonation was the most commonly cited AI-enabled attack type, appearing in 45 percent of responses. AI-assisted phishing and malware activity are also increasing, enabling criminals to produce convincing messages, automate reconnaissance, and adapt campaigns more quickly.
A Practical Enterprise Response
- Create and maintain an inventory of approved AI models, agents, plugins, and data connections.
- Apply least-privilege access to every AI service and non-human identity.
- Review cloud configurations and API permissions continuously.
- Require additional verification for financial requests, password resets, and sensitive instructions delivered through voice or video.
- Record agent actions so investigators can reconstruct decisions and data access after an incident.
- Include AI services in incident response, backup, and breach-notification plans.
IBM also found signs of growing resilience. Sixty-one percent of surveyed UK firms plan to increase cybersecurity investment following a breach, with incident response, AI governance, and data protection among the priorities.
In my view, organizations should avoid interpreting every incident involving an AI application as evidence of an entirely new class of attack. Many failures still begin with exposed credentials, excessive permissions, insecure APIs, and incorrect cloud settings. The difference is that AI systems can connect these weaknesses at greater speed and scale. Effective AI security therefore begins with strong identity, data, cloud, and software governance rather than a separate collection of controls built around the AI label.