AI Is Not Slowing Down. Has Your Risk Register Caught Up?

What Are the Threats Relating to AI? 

IBM reports that there has been a 56% increase in AI-driven attacks, including deep-fake and AI-enabled malware, with 92% of organisations that reported an AI-related breach lacking proper AI access controls.

The National Cyber Security Centre (NCSC) has also reported that cyber attacks will increase in volume and sophistication as AI’s growing cyber capabilities make it easier and faster for attacks to be conducted. In its Annual Review 2026, the NCSC states that AI makes cyber attacks “easier, faster and cheaper to execute, and at greater scale”.

Over the past year, the NCSC's Incident Management team also received 1,831 reports of cyber attacks, 80 of which were nationally significant. By 2028 the NCSC expects that AI-powered malware will likely be used in a ransomware attack on UK critical national infrastructure. 

The threat is credible and the impact to your organisation should be evaluated through a risk assessment. Thankfully, the rules of the game haven’t changed. Risks should be evaluated, prioitised and mitigations put in place to ensure your information security programme is staying up to date with the threat landscape. 

Risks Worth Considering 

Here are five risks to consider evaluating and adding into your information security risk register: 

Data Leakage Via Shadow AI

AI tools and services are available via the web and within applications. It is increasingly difficult to control access to these tools and services. Well-meaning employees could submit confidential information into unapproved tools, which could lead to the unauthorised disclosure or transfer of data. Incidents involving shadow AI have more than doubled to 43% this year from 20% last year. 

AI-enabled Social Engineering

Criminals are using AI-generated voice, video and email to trick employees into sharing information or making payments. In 2024, a North Korean actor was recruited by KnowBe4 using a stolen US identity and was sent a company laptop. This was despite KnowBe4 carrying out four video interviews, background checks and ID verification with the actor.  

Attacks Against Your AI Systems

The highest risk to large language models (LLMs) identified by OWASP is prompt injection. Malicious instructions may be hidden which are later processed by the AI model which could lead to the disclosure of confidential information or the executing commands in connected systems. Additional attack techniques include model poisoning and improper output handling.  

Excessive Access for AI Assistants, Agents and Integrations 

Ungoverned AI assistants, agents or connections such as APIs or MCPs may have excessive access rights that could lead to the unauthorised loss, modification or deletion of data. 92% of organisations that reported an AI-related breach lacked proper AI access controls according to IBM’s Cost of a Data Breach Report 2026. The NCSC warns that “a single failure can become a serious incident” as the autonomy and complexity of agentic systems can increase the attack surface and make behaviour harder to predict. 

AI Supply Chain Risk 

AI assistants and agents are susceptible to vulnerabilities including traditional vulnerabilities such as outdated or depreciated components, as well as vulnerable third-party pre-trained models. These vulnerabilities could be exploited to compromise the AI system or tamper with the models, compromising the integrity of outputs or being used to introduce malicious code. 

Risk Mitigation Strategies 

The global average cost of a data breach is around $5 million. Here are some of the most important risk mitigation strategies to consider to prevent or reduce the potential cost to your organisation: 

• Policies and standards – Establish policies, requirements and guidelines for development, deployment, acquisition and use and maintenance of AI systems within your organisation. 
• AI Inventory – Build up a comprehensive inventory of all AI assistants, agents, integrations, models and data sets in use across your organisation. This should include security-relevant information such as the sources of training data, guardrails and cryptographic material. 
• Securing AI systems – Whether you develop AI systems from scratch, deploy or integrate AI technologies, secure by design principles should be embedded throughout the system’s lifecycle. The NCSC has established Guidelines for Secure AI System Development.  
• Restrict access – Classify your information and enforce the principle of least privilege and review permissions prior to connecting assistants or agents, this includes API keys, service accounts and MCPs.
• AI-enabled cyber defences – Identify opportunities to use AI as part of your information security programme, particularly within defensive operations such as security monitoring. 
• AI-Data Leakage Monitoring – Deploy technologies that can help detect and block potential AI-related data leakage channels, including DLP and CASB, and monitor the dark web for data breaches. 
• Resilience to AI-enabled attacks – Plan and prepare for AI-enabled attacks that may impact your organisation, considering realistic scenarios. Establish playbooks and rehearse your response. 
• Educating and awareness – Raise awareness of the current threats, how to spot them and what employees in your organisation should do to report it.
• Watch the frontier – Continue to monitor the threats and technological developments of AI to ensure your information security programme can stay ahead of the risks.