AI agent security assessment
Available nowFind out what your AI coding agents can reach.
Your people are running Claude Code, Codex, Cursor, Copilot, or agents they wrote themselves, including on machines your MDM doesn't manage. Each agent can read files, run commands, hold API keys and reach the internet. We find out what they can do on your network, then write a policy for them and put the controls in place.
About one week · Up to about 25 machines · Windows, macOS and Linux · Price agreed before work starts
Who it's for
Machines your MDM and endpoint security don't cover.
Linux workstations and servers, research and GPU machines, developer laptops, contractors' machines: the parts of a Windows, macOS and Linux fleet that often sit outside device management.
If you can't say which AI agents run on which machines, or what they can reach, this is for you.
The assessment
Four deliverables, in about one week.
- 01InventoryThe AI agents and automations on each machine, including ones your team wrote that call a known model provider or a local model server; how each one was started (by a person, a remote session, a scheduler, a service, or another agent); and which model providers it talks to. Collected read-only.
- 02Configuration reviewThe MCP servers each agent can use, where API keys and tokens sit in config files, and which agents approve their own actions.
- 03Network reviewWhat the agents can reach outbound, and the firewall and egress rules that would contain them.
- 04A written reportFindings ranked by severity, the fix for each one, and a one-page summary for leadership.
How an engagement runs
From the first call to a team that runs it.
- 1. Onboarding callWhat you run, on which machines, and what worries you. We agree the scope and the price.
- 2. AssessmentAbout one week: inventory, configuration and network review, and a written report.
- 3. Policy and enforcement planA written policy for AI agents in your organization, and how each rule is enforced on each kind of machine: which agents are allowed, and what they may reach.
- 4. ImplementationPutting the controls in place: firewall and egress rules, network segmentation and access control, the agents' own managed settings where the tool offers them, and the bastion where it fits.
- 5. Turnover and trainingYour team gets the runbooks and a training session, so they can keep the policy enforced and monitored after the engagement ends.
The assessment is the first step. Each later step is quoted from what the assessment finds, so you know the cost before that work starts.
The tool
Open source, so you can read what runs on your machines.
The bastion is a single program for Linux, macOS and Windows. It finds AI agents by their processes and their network traffic, and records how each one was started. Its source is public under the GPL.
Who does the work
I'm Casey.
I've spent 16 years in network infrastructure, security, and operations: as a network engineer for the US Air Force, ITA Pentagon, and lead network access control engineer at a federal research campus.
I started Privatae when I saw the systemic issues created by bringing AI into a network that isn't ready for it.
Book a call
Tell us what you run.
Which tools your team runs and how many machines are in scope. We reply by email with dates for the first call.