Securing LLMs
AI security and governance for Michigan businesses
Let staff use AI assistants without handing them the file server.
Securing large language models in a business: which tools are allowed, what data may enter them, how people sign in, what an agent may do on a computer, and the logs that prove it. Covers Claude, Microsoft 365 Copilot, ChatGPT and the agents built on them.
What's included
Treat the assistant like a new hire with a very fast computer.
AI tool inventory & shadow AI
Which assistants staff already use, found from DNS logs, browser extensions and expense reports. Then a decision on which are allowed, and a block on the rest at the firewall and in the browser.
Acceptable-use policy
One page that says what may go into a public AI tool, what needs the company tenant, and what never leaves. Written with you, acknowledged by staff, and enforced by the controls on this page rather than by hope.
Enterprise tenant & SSO
The business plan for each approved tool, single sign-on through Microsoft Entra, MFA, and provisioning tied to HR, so a departed employee loses the assistant along with everything else.
Data controls for prompts & files
Training opt-outs confirmed in writing, retention settings set, DLP rules that stop card numbers and patient records entering a prompt, and sensitivity labels that Copilot respects.
Permissions before rollout
An assistant answers from everything the signed-in user can open. Inherited permissions and old shares get cleaned up before the assistant is switched on, not after it surfaces the payroll folder.
Connector & MCP review
Every integration an assistant can call, from email and files to the CRM and accounting system, reviewed for who authorized it, what it can read, and whether it can write. Unapproved connectors blocked by admin policy.
Agent & desktop sandbox controls
Agents that run commands on a workstation, such as Claude Cowork and coding assistants, configured through MDM: network allowlists, folder access limited to the job, and the sandbox's own security sensor where the vendor supports one.
EDR inside AI sandboxes
Claude Desktop can load your EDR vendor's signed sensor into the Cowork virtual machine each time it starts, deployed by MDM or registry policy. We confirm your vendor is approved, package and deploy the sensor, and verify its alerts reach the SOC.
Logging & audit evidence
Audit logs from each AI tenant exported to your SIEM, a monthly report of who uses which tool, and the written answers for the insurer or auditor who asks what your AI policy is.
In detail
Why AI tools need their own security page
An AI assistant is a new kind of user. It reads whatever the signed-in person can read, it talks to a vendor's cloud, and increasingly it acts: sends the email, edits the file, runs the command. The controls on our cybersecurity page cover the device and the identity. They do not, on their own, decide which assistant staff may use, what data may be pasted into it, or what an agent may do on a laptop. That is this page. It applies whether the tool is Claude, Microsoft 365 Copilot, ChatGPT, or an assistant a vendor added to software you already own.
Worked example: EDR inside Claude Cowork
Claude Cowork runs commands in a virtual machine on the user's computer, and the EDR sensor already on that computer cannot see inside it. Anthropic's answer, in beta for Team and Enterprise plans on Claude Desktop for macOS and Windows, lets an organization name an approved EDR vendor and push that vendor's signed sensor package by MDM or registry policy. Each time the virtual machine starts, Claude Desktop checks the package signature and runs the vendor's installer inside it. The sensor then reports to the vendor's cloud like any other endpoint. CrowdStrike, SentinelOne, Palo Alto Networks and Tanium are among the approved vendors. The fine print matters: the sensor can see prompts, files and responses; Cowork sessions that run on Anthropic's servers are out of scope; and removing the policy stops future installs but does not uninstall a sensor already in place. We read that fine print so your team does not have to.
What an engagement looks like
It starts with a short review: which AI tools are in use, which plans they are on, how people sign in, what the tenant settings say about training and retention, and which connectors have been authorized. The review ends in a ranked list with costs. Then the policy and tenant settings go in, sign-on and provisioning are tied to your identity provider, and the agent and sandbox controls are deployed by MDM. From then on, AI usage and alerts appear in the same monthly report as the rest of your security program. This can be bought with managed IT, added to your own IT team's work, or layered on another provider's support.
Who this is for
Businesses whose staff already use AI tools, leaders asked to approve Copilot or Claude for the whole company, and practices that handle regulated data.
Every industry we serve pulls on this area differently. Jump to yours to see what it usually looks like.
How engagements run
No mystery scope, no surprise change orders.
- 01
Walk it
We look at the space, the stack, or the drawing set before quoting anything.
- 02
Scope it
A fixed written scope: what is included, what is not, who does what, and when.
- 03
Run it
One project manager owns it end to end and stays your contact after handoff.
Related services
Usually bought together, or one after the other.
Where we work from
AI security from Troy and Pontiac, Michigan.
Remote support covers any location. On-site work across Metro Detroit is scheduled from the nearer office.
Should we just block ChatGPT and Claude?
Blocking public tools without offering an approved one pushes the work onto personal phones, where you have no control at all. The usual answer is a business plan with single sign-on for the tools you approve, a short policy on what may go into them, and a block on the rest.
What does the Claude Cowork EDR sensor actually do?
Each time the Cowork virtual machine starts, Claude Desktop copies your EDR vendor's signed sensor package into it, checks the signature against the vendor's key, and runs the installer. The sensor then reports activity inside the VM to the vendor's cloud. It is in beta for Team and Enterprise plans and is configured by MDM on macOS or registry policy on Windows. Anthropic's support article lists the approved vendors and requirements.
Does our existing EDR already cover AI tools?
It sees the browser and the desktop app as processes. It does not see inside a sandbox or virtual machine an agent uses to run commands, and it does not see what data goes into a prompt. Those need the sandbox sensor, DLP rules and tenant audit logs described on this page.
Is this the same as your AI & Automation service?
No, but they fit together. AI & Automation decides what to automate and cleans up permissions first. This page secures the tools themselves: sign-in, data handling, connectors, agents and logging. Most clients do the readiness review and the security review together.
Which frameworks do you map AI controls to?
The NIST AI Risk Management Framework for the governance language, and the same CIS Controls v8.1 and NIST CSF 2.0 baseline as our cybersecurity page for the technical controls. We do not certify compliance. We run the controls and keep the evidence your auditor or insurer asks for.
Tell us what's in front of you.
A drawing set, a failing switch, an audit deadline, or a hunch that IT costs too much. We'll tell you who picks it up and what it takes.