2 min readBy Rexaris AI
Claude Code in a regulated company: what security needs before developers get access
Coding agents are the fastest-growing AI workload in enterprises. Here is the short list of controls that let security say yes.
Developers want Claude Code. Security teams in banks and payment companies want to say yes, but they have a specific set of worries, and "trust the vendor" does not answer any of them. This post lists what a security lead typically needs in place before a coding agent is approved, and what each item looks like in practice.
1. One corporate endpoint, no personal keys
The first problem is not data, it is keys. A personal Anthropic key in a developer's shell is unmanaged spend and an unmanaged data path. The fix is a corporate gateway that speaks the Anthropic API: Claude Code is pointed at it through an environment variable, and every request carries the developer's identity rather than a shared secret.
2. Secrets never leave the perimeter
Coding agents read files. Files contain .env entries, cloud credentials, private keys and tokens. A secrets detector on the gateway catches the common formats (AWS access keys, GitHub tokens, JWTs, private key blocks) and applies a hard rule:
Data class = CREDENTIALS → BLOCK
The request is not forwarded, the event is logged, the developer gets a clear message. There is no "mask" option for credentials; a masked key is still a leaked shape.
3. Source code policy you can explain
Whether proprietary code may go to an external model is a business decision, not a technical one. What the gateway provides is the ability to express the decision as a rule per repository, team or data class, and to prove afterwards that it was applied. Some organisations allow it for most repositories and route a few sensitive ones to an internal model. Others start with audit-only and decide later.
4. Usage and cost per developer
Agents consume far more tokens than chat. A budget per developer, project or team, with alerts before the limit, turns an unpredictable bill into a line item. It also surfaces the one script somebody left looping over a repository overnight.
5. An audit trail that answers questions
Who used which model, from which tool, with which policy outcome, at what cost. When an auditor asks, the answer should be a query, not an investigation.
What is still ahead
Tool and MCP governance for agents is an area we treat as roadmap, not as a shipped control. Agents calling tools that modify systems need approval workflows, and that work is in progress. Until then, treat agent tool access as a separate review item rather than assuming the gateway covers it.