Artificial intelligence is rapidly evolving from a productivity tool into part of the operational fabric of businesses. Companies are using AI to analyse data, communicate with customers, detect fraud and automate processes. Increasingly, AI agents can interact with other systems and complete tasks with limited human intervention.
For businesses, the question is not whether to adopt AI, but how to give it greater autonomy safely. This is where AI security testing and continuous risk monitoring can become valuable tools that allow AI to go further within business workflows.
The real value of AI security testing
Every organisation wants to capture more value from AI. As systems move from drafting and analysis into customer service, decision-making and transactional processes, however, businesses need confidence that their controls will continue to work.
AI security testing, including red teaming, deliberately challenges an AI application to identify vulnerabilities before they result in an incident. Testing can examine prompt injection, guardrail bypass, disclosure of confidential information, multilingual attacks, excessive permissions and unintended actions by AI agents.
Where a system reads images or documents, testing extends to what is hidden inside them, such as instructions concealed in a PDF, spreadsheet or image that the model then acts on. The reverse problem matters just as much: whether a model can be relied on to spot a manipulated identity document, a synthetic face image or a fabricated CV.
Rather than relying solely on policies or assumptions, businesses can generate evidence of how their AI behaves under pressure.
The output is evidence rather than assurance. Testers generally report a headline attack success rate, being the proportion of attempted exploits that worked, alongside severity mapping and a remediation list. Vulnerabilities can then be remediated and the system tested again.
The result is a safer AI application and greater management confidence to allow AI to perform more valuable and increasingly autonomous tasks.
From AI assistants to autonomous workflows
The risk changes as AI gains responsibility.
An AI assistant drafting an internal document presents a different exposure from an agent accessing customer information, interacting with business systems or initiating a transaction.
The objective should be to introduce autonomy progressively, with controls proportionate to what the AI is permitted to do.
Continuous monitoring is important because AI systems are not static. Models, integrations and permissions change. Testing and monitoring can help businesses understand whether those changes have introduced new vulnerabilities and provide assurance as AI is expanded into higher-value workflows.
MIT’s AI Incident Tracker reinforces why this matters. Of the real-world incidents it has classified, almost all occurred after deployment, with only around 2% arising before release, and roughly half were recorded as intentionally caused. This is an operational exposure, driven substantially by deliberate misuse, and it calls for operational evidence.
Source: MIT AI Risk Initiative, AI Incident Tracker (https://airisk.mit.edu/ai-incident-tracker/incident-overview)
How insurers are assessing AI risk
Insurers are beginning to face the same question: how can they differentiate between businesses using AI responsibly and those where the exposure is poorly understood?
AI underwriting may follow a similar path to cyber insurance. Cyber underwriters increasingly look for evidence of controls such as multi-factor authentication, endpoint detection and backups. For AI, insurers may consider what systems are deployed, what data they access, whether they can take autonomous actions, where human approval is required and whether applications have been independently tested.
AI security testing can provide useful underwriting evidence. Rather than simply telling an insurer that guardrails exist, a company may be able to demonstrate that those controls have been challenged, weaknesses identified and improvements made.
How can AI risk be insured?
There is not necessarily one policy covering every AI exposure. Depending on how AI is used and the circumstances of a loss, relevant coverage may sit across Cyber, Professional Indemnity or Errors & Omissions, Directors & Officers and Crime insurance. Emerging affirmative AI insurance solutions may also become increasingly relevant.
Coverage will always depend on the individual policy wording and facts of a claim. A significant amount of AI exposure currently sits in wordings drafted before AI was contemplated, so businesses should review how their existing programme responds as their use of AI evolves.
AI security as an enabler
Although speed of deployment is a factor, the businesses that gain the greatest value from AI may be those that can demonstrate enough control and visibility to trust AI with increasingly important processes.
Security testing and continuous monitoring can help create that confidence. They allow businesses to identify weaknesses, improve controls and demonstrate to insurers that AI risk is being actively managed.
For businesses, this can turn AI security from a defensive exercise into an enabler of greater automation and innovation.
At Elmore, we work with businesses and insurers to understand how emerging technologies are changing risk and how insurance can evolve alongside increasingly autonomous AI