The impact of AI on the speed and scale of cybercrime seems certain to be one of the dominant security themes of 2026. As the World Economic Forum (WEF) pointed out in its recent analysis, “generative AI is boosting the speed at which hackers work, from identifying new security weaknesses to writing the matching malware.”
If that wasn’t enough, recent headlines about agentic frontier AI models from Anthropic and OpenAI autonomously instigating breaches have added another item to the list of emerging security priorities. And we already know that threat actors are using agents to automate research and produce convincing personalized phishing or social-engineering messages for likely targets.
Then there’s the massive adoption of AI tools and agents on an operational level. These systems are typically connected to business applications and sensitive data. When they are granted excessive permissions, attackers may be able to manipulate or misuse them to act on their behalf.
The point is, security teams are in uncharted territory, and as the WEF also explains, “the window for [them] to detect and respond is becoming ever smaller, and they are struggling to keep pace.”
Familiar weaknesses
Behind the growing sense of alarm, however, the stages of an attack have not fundamentally changed. Take end-user devices, for example, which remain a key target because compromising just one can give an attacker a foothold inside the organization.
If we follow the logical chain of events, the damage an attacker can cause will depend in part on the permissions available to the compromised device or account. AI can also accelerate the activities that follow the initial breach by helping attackers gather information and identify likely targets more quickly.
The same level of concern applies when an AI agent with broad access to internal systems and data is manipulated to act on an attacker’s behalf. Security teams need to understand what those systems can access and maintain a clear record of the actions they take. In a word, this is all about visibility.
The basis for effective visibility is that organizations must have a clear picture of activity across their technology environment. That might sound obvious, but it can be very difficult to achieve when systems are constantly changing, and AI agents act at machine speed. Indeed, agent visibility is already a significant problem, with research from the Cloud Security Alliance revealing that “only 21% of organizations maintain a real-time registry or inventory of their agents.”
When it comes to user endpoints, the picture is no less concerning. A recent survey of 500 IT leaders found that 60% of organizations have devices they cannot fully manage. It also identified “full visibility into every device from a single platform” as respondents’ number-one goal for the year ahead.
Monitoring what matters
So, what needs to change to bridge this yawning visibility gap? It sounds obvious, but success depends on collecting reliable activity data from the devices employees use every day. Yet, as we have seen, this is far from a given and is especially difficult in technology environments characterized by constant change.
That’s not to say organizations haven’t tried to monitor the status and security of their endpoints, but the problem has always been maintaining a complete and current picture as the endpoint estate changes. Periodic checks, which may run only every few hours or once a day, may not give security teams enough information to understand what is happening as it happens.
Instead, endpoint monitoring needs to provide more timely data on significant changes to a device. This needs to be combined with audit logging that records actions taken within the software and services connected to the organization’s environment. Together, these records give security teams the evidence needed to investigate unusual behaviour, which is particularly important when an AI agent can access sensitive data or take action without waiting for a person to intervene.
Crucially, visibility is only useful if security teams can identify who, or what, was responsible for an action. AI agents should therefore have separate, non-human identities rather than share user credentials, allowing IT and security teams to distinguish their activity from that of employees. If suspicious behaviour is detected, the organization can revoke the relevant agent’s access without disrupting human users.
Separate agent identities also provide the basis for assigning task-specific permissions, so each agent has only the access it needs to carry out its intended role. This helps limit damage if an agent is manipulated into acting on an attacker’s behalf.
One way to manage higher-risk AI changes is through configuration-as-code, which creates a documented record of each proposed change and supports review before deployment. If something goes wrong, teams can return to a known-good state, which offers important security and operational advantages.
Join our LinkedIn group Information Security Community!
