I'm the IT manager at a small technology company, and leadership and the development team are moving quickly toward using Claude throughout the business. I understand the productivity benefits, but I'm responsible for protecting company data, infrastructure, credentials, and user access, so I want to understand the risks before this becomes widespread.
My biggest concern is that I don't yet know enough about how Claude, Claude Code, Cowork, connectors, MCP servers, and related integrations interact with our systems. I'm looking for formal administrator or security-focused training, along with practical guidance from people who have managed a rollout. I want to establish sensible controls without unnecessarily preventing employees from doing their jobs.
5 Answers
Use a Team or Enterprise plan rather than personal accounts so you get organizational controls and contractual protections around company data. Start with restrictive defaults: disable Cowork, connectors, remote-control features, and other integrations until they’ve been reviewed. Enable capabilities gradually as you understand the risks, and use groups or roles to limit access. An enterprise tier can add audit logs, custom roles, stronger provisioning, and usage controls, but it can also become expensive because of metered usage.
Create a formal AI acceptable-use policy and an approval process for integrations, connectors, and MCP servers. Train staff not to put confidential information, secrets, regulated data, or customer material into personal AI accounts. Work with leadership to document the company’s risk appetite, then provide an approved, controlled option instead of trying to block everything.
Also measure actual usage and spend. Group-level limits, per-user budgets, alerts for unusual consumption, and periodic license reviews help prevent runaway costs and make it easier to identify risky behavior.
Anthropic’s administrator and user courses are a good starting point. I’d also review its Trust Portal, security documentation, and material on securing agentic coding tools. The product-specific training helps, but you should combine it with broader subjects such as DLP, secrets management, identity, least privilege, and agent security. OWASP’s guidance for generative and agentic AI and the NIST AI Risk Management Framework are useful references too.
Treat Claude as a privileged automation identity, not merely another chat application. Inventory who is using it, which machines it runs on, what files it can access, which MCP servers and connectors are enabled, and which API tokens or credentials are exposed. Use narrowly scoped, short-lived, revocable credentials and log activity and data egress.
DLP can help catch users pasting sensitive code or data into prompts, but it may not see an agent writing files or running shell commands locally. For coding tools, consider sandboxing, approval gates for command execution, tool-call logging, and endpoint management. The surrounding permissions and integrations are usually a bigger risk than the chat interface itself.
If you’re heavily invested in Microsoft 365, Copilot may be worth evaluating because it offers familiar identity, compliance, and data-governance controls and can provide access to Claude models in some configurations. That doesn’t make it automatically secure, though. Poor sharing permissions and weak user training can still expose sensitive information, so the underlying data-access model and user education remain essential.
Exactly. A managed platform can improve governance and accountability, but it can’t compensate for misconfigured sharing or users putting sensitive information into the wrong tool.
Thanks for collecting these resources. I had found some of Anthropic’s training, but I was specifically hoping to hear which materials were useful from people who had actually gone through a rollout.