AI is only as secure as its access

Artificial intelligence is increasingly integrating in the communications and collaboration platforms that defence and national security teams rely on every day. And it possesses many benefits, such as for analysts who must work through volumes of information no person could read in the time available, and commanders who need to decide faster than the people they are up against. AI promises quicker work, sharper judgement and fewer mistakes, and it can genuinely deliver all three. Used well, it turns a flood of data into something a person can act on and frees skilled people for the work that actually needs them.

However, a secure communications platform is not a consumer messaging app with a helpful assistant bolted on. Its entire value depends on the right information reaching the right person, and travels no further. The moment you introduce an agent that can read across conversations, retrieve from connected data and act on behalf of a user, you have added a new participant. The question then becomes what the AI technology is allowed to see, and how it accesses it.

Who controls the data

In these settings, access has always been governed by identity, entitlements and the need-to-know. An AI agent does not inherit that discipline by default. If an agent can reach everything a platform holds, it can surface everything, including to people who were never meant to see it. Oversharing by a machine is still oversharing, and it happens faster and at greater scale than a person ever could. So the first governance question any organisation should ask is does the agent control the data, or do people?

In practice, an agent acting for a cleared user is not simply that user. It can move through in seconds what would take a person days, and it can draw together fragments that were harmless on their own and revealing once combined. Restricting it is not only about which systems it may enter, but how much it can gather in one place and how much it is allowed to pass on. Leakage becomes something to design against from the outset rather than to notice afterwards. Access has to stay correct whether the request comes from a person or from software acting in their name, and it grows harder and more important once agents are the ones asking.

Being present is not being safe

It is tempting to assume that once AI is in place the hard part is over. But Governance is correctly fast becoming a discipline in its own right, built on knowing where data came from, enforcing policy on how it is used, and monitoring what an agent has touched and why. Training is where much of the risk is settled long before anyone types a question. A model grounded on material gathered without the same care over classification and permission can reproduce it in places it was never meant to appear, so trust in an answer depends on being able to trace where it came from and why the agent was entitled to use it. Human oversight remains the fixed point. AI can accelerate a decision and improve it, but it cannot be the thing that owns it. The person accountable for the outcome still has to be in the room.

Zero Trust has to reach the agent

The sharpest change AI forces on us is the need to extend zero trust to actors that are not human. For years, zero trust has meant refusing to assume trust and verifying every user and every request. Agents now need the same treatment.

They need identities of their own, they must be authenticated and entitled to the data they reach, and they must be watched continuously once at work. Chief Information Officer Douglas Cossa put it well, that we will need a common identity system spanning humans and AI. In practice, that means treating an agent much as you would a user or a service, as something that must prove it is permitted to be where it is, every time. It also means taking seriously the tools that slip into an organisation without oversight, adopted quietly by teams keen to move faster, because those are the ones sitting outside the identity and monitoring model. AI operating outside that model is an open door.

The value of AI in this domain does not come from raw capability. It comes from AI that is governed, supervised by people and held to the same access rules as everyone else, working on trusted data to help teams decide well and act quickly. The benefit is real and it is arriving now. But speed counts for nothing if the machine can reach what it was never meant to touch, and trust that cannot be enforced is not trust at all.