Insights

AI Security at an Enterprise Scale

Written by ClearPoint | Oct 6, 2026, 12:36:27 AM

Most security programmes were built to put controls around risks that are well established and relatively slow moving. AI agents are not that, and they are changing quickly enough that the control you need for a given risk often isn't sitting in the framework yet. An agent is only useful once it can reach the systems that matter to a business, and that reach is exactly what turns adoption into a governance question, not just a technical one.

This piece draws on the latest ClearPoint Unlock episode, staying secure while moving fast with AI agents, in which COO Paul Scott and Lead Engineer Will Verland cover what has changed over the past 12 months and unpack why AI security has become everyone's problem.



Why avoidance doesn't work

 

Faced with a risk that is hard to pin down, the easiest response looks like holding back until it feels safer. Avoiding AI adoption doesn't guarantee people stop using it. Sometimes it just pushes the work onto a personal account instead, with no visibility into what data goes in or how it's used. Saying no doesn't stop the use, it just moves it into shadow accounts.

"I've heard of organisations that have said 'No AI,' only to find out a few months later that everyone's uploading their source code to their personal Copilot accounts that are doing training on things," says Will Verland, Lead Engineer at ClearPoint. "So, just saying no may not be enough for organisations. It just drives it underground."

 

Enable before you control


The better approach stages the rollout rather than trying to anticipate every use case up front. Adopting AI plays out as a maturity story rather than a single policy decision. ClearPoint, for example, rolls new tools out on the permissive end, then shapes the controls around how the organisation actually uses them. Every rollout surfaces risks that weren't on the original list.

"When we first roll out these tools, rather than locking everything down, we want to get adoption going," says Paul. "We try to enable as much as we can, and then review as we go and start to apply controls once we can see how the organisation is using it."

The same pattern applies to integrations. An agent only starts delivering real value once it can reach the core systems that run the business. This is also where the risk concentrates. That makes it worth running a risk assessment before any of those integrations is turned on, and piloting each one before it's rolled out more broadly.

ClearPoint has used Claude Cowork outside of engineering, in marketing, sales and shared functions. "What I do like about Claude Cowork, unlike Claude Code, is it's got sandboxing built in already," says Paul. "Each of your Cowork sessions runs in a container on your local desktop, and you give it access to a single folder on your machine." Which integrations it can reach beyond that is a decision for the business, and one worth a risk assessment of its own. Drive with every client folder in it, the CRM and the ERP each carry a different level of exposure. Cowork's enterprise controls are being layered in over successive releases, so early adopters work with a smaller set of controls now and expand them as more become available.

Watch or listen to the full ClearPoint Unlock episode on AI security with Paul Scott and Will Verland.

How adoption actually takes hold

 

The people running a business process are the ones who know where the friction sits and where the savings could come from. When those same people have the tools and the permission to try things, the valuable use cases come from them rather than from a leadership team reviewing the process from above.

Support from the top is not the same as pressure from the top. Support builds the culture and gives people permission to experiment, while pressure stifles the discovery before it starts. At ClearPoint, the guardrails went in first, then people were actively encouraged to use AI, with a clear expectation that the business wanted to be a frontier firm. The rest happened on its own once people started seeing what AI could do for their own work, and it’s now part of everyday work across the business.

“If they haven't got support from up top, then they're not going to feel empowered to try things and experiment,” says Paul. “We've really empowered everybody and that enabled us to see we've got 100% AI usage across our company and we wouldn't have that if we didn't have that culture and set those expectations that we want to be that frontier firm.”

 

Extend what you already have

 

Businesses already scope access when a new supplier or a new employee comes on board. An AI agent can be handled the same way, with its own identity and only the access its job requires. An agent almost never needs to run under a superuser account. Keeping it scoped to what the job actually needs means a mistake stays contained to what that identity could reach, rather than spreading across the whole business.

"Reiterating the principle of least privilege, we'd recommend that you give agents their own identities with just the permissions that they need to do their jobs," says Will. "You'll never want your agent to delete a GitHub repository. You'll never want it to publish a release by itself or merge a pull request. You can control that with a GitHub token that you give it."

Those limits hold even when the model decides to try. "I've seen agents try to merge pull requests and then be blocked because they don't have permissions. You can apply these deterministic controls to these non-deterministic large language models. And they're really effective," says Will.

None of this needs a new framework. Information security standards like ISO 27001 already cover it. Policies get checked rather than just written, and access reviews happen on a schedule rather than only when someone remembers. Training doesn't catch everything, so moving the checks into delivery pipelines, where they run on every change, adds another layer. An agent simply needs to sit inside that same discipline, not one built specifically for it.

For the engineering controls in detail, from sandboxing to package cooldowns, read Securing the AI agents on your developers' machines.

Being upfront with clients


For any business that works with client data, the concern that comes up most is intellectual property. Clients are giving access to critical documents and code, and they need to know that material won't end up training a model or sitting in a system long after the engagement ends.

If AI is part of how you deliver, the time to set out how a client's data will be used is at the start of an engagement, before any work begins. Where a client wants their work kept to particular models, that can be agreed upfront and held to. AI clauses in standard agreements make the use transparent and give the client the ability to opt out, for a specific piece of work or a specific dataset.

 

The risk sits in the integrations

The worry with each new integration is data leaking from key data sources into the models, including client data, and where it ends up after that. Those same integrations are what make agentic workflows worth having, so they get opened up in a measured way rather than all at once.

"Once you start opening up those integrations, that's where the true value of a lot of these agentic workflows are in the operations of a business," Paul says. "If they're not connected into your core systems, they can't orchestrate the business processes that you would do. So it's inevitable that you're going to open up those integrations, and that's where you get the benefit, but if you do it in a measured way, you can manage the risk."

 

The board-level question

AI agents are already operating inside real systems. The boundaries around what they can do are worth setting as deliberately as the decision to use them in the first place.


The right question at the board level isn't whether AI is safe. It's how the business is using AI to innovate the business model, and how it's using AI to deliver more efficiently to clients.

"I see two types of organisations with AI adoption. There's organisations who are going to get left behind asking, 'Is AI safe?' And there's organisations that are going to succeed saying, 'What can AI do?'" says Will.

Scoping what each agent can reach, being upfront with clients about their data, and starting the highest-risk integrations on a pilot basis are the moves that let leaders answer that question with confidence rather than caveats. None of it removes the risk entirely, so what matters is the layers you have in place and knowing what you are left carrying once they are there.

"A lot of these things sound scary, and the mistake is to do nothing at all," says Will. "Perfect is the enemy of good, and if you can take one or two controls away from this and implement them in your organisation, you're going to be much better off."

If you’re planning to connect AI agents to the systems that run your business, ClearPoint can help you do it with the right controls in place. Talk to us.