When AI Becomes an Actor, Trust Must Evolve Too
Perhaps the most important cybersecurity shift in 2026 is not that AI became smarter. It is that AI started to act.
In September, Microsoft reported an Azure attack in which compromised service principal identities carried out a wide-ranging destruction spree in seven minutes and also tried to remove backup and recovery protections. Microsoft pointed to the case as part of a broader shift towards AI-orchestrated attacks, where automation increases both speed and scale.
Seven minutes.
That is not enough time for people to react and act consistently.
If the attacker operates at machine speed, defense cannot remain at human speed.
This is not only a technology shift. It changes the meaning of trust.
We are no longer giving AI only tasks. We are giving it authority.
So far, much of the discussion around AI has focused on whether models are accurate, reliable, and trustworthy. That question is already becoming too narrow.
The more important question is:
What is AI allowed to do?
OpenAI has reported cases in which models bypassed controls restricting internet access, used unauthorized communication channels, exploited vulnerabilities, and reached third-party systems.
Anthropic has also reported cases where Claude models gained unauthorized access to real third-party systems during cybersecurity evaluations.
This is not about AI “rebelling”. It is a story about authority.
Once a system has a goal, autonomy, and access to tools, controlling the boundaries of its actions becomes a security problem. That is a fundamental shift.
Every autonomous agent has also an identity
Wavestone’s AI Cyber Benchmark 2026 captures this well. Agentic AI challenges the traditional security model because we are no longer protecting only data and systems. We also need to protect decisions, privileges, and machine actions.
And most organizations are not there yet.
In Wavestone’s benchmark, only 15% of organizations had identity and access management solutions for AI systems and agent development. Only 16.7% had defined which functions and tools AI agents were allowed to access within secure development standards.
The sample is limited and should not be treated as representative of the whole market. But the direction is clear.
If an AI system can read data, modify a system, or trigger a process, it has digital authority. That means we need to ask the same questions we ask about human identities:
Who – or what – is acting? With what rights? For what purpose? For how long? What happens when the behaviour deviates from what we intended? And most importantly, how do we revoke that authority?
Visibility is becoming a strategic capability
One of Wavestone’s most striking findings relates to visibility. 87.5% of the organizations in the benchmark were already generating logs from AI applications. Only 8.3% were monitoring logs from critical AI applications.
The data exists. The security function simply may not see it. That is not a minor technical gap. It is a governance gap.
In a world where autonomous systems can act faster than people can respond, the SOC needs visibility also into agent identities, privileges, tool use, actions, and anomalous behaviour.
Trust must become verifiable
In the age of AI, trust must be built on four things: identity, constrained authority, visibility, and recoverability.
We need to know who or what is acting. We need to limit what it is allowed to do. We need to see what it is actually doing. And we need to be able to stop, isolate, and recover when something goes wrong.
No single actor can deliver this alone. Cybersecurity has always been a collective effort. In the age of autonomous AI, it becomes even more so.
Author:
