AI Security · Coding assistants
AI coding assistants can accelerate software development, but developer workflows may also involve source code, configuration files, credentials, and internal technical documentation. GuardStance protects proprietary repositories, API keys, and customer data across Claude Code, Codex, GitHub Copilot, and Cursor—without slowing down your developers.
Book a demo
core-banking/ Red · critical The assistant cannot read it Block nothing leaves payments-api/ Yellow · internal Readable, through the gateway Record secrets masked docs-tools/ Green · open Readable Monitor logged A developer asks an AI assistant to help troubleshoot an application. The request may include source code, configuration values, or internal technical details that should not be shared externally.
GuardStance enforces repository-level access rules and inline secret masking before any prompt or file context leaves the developer's machine.
Claude Code and Codex, through managed hooks.
Anthropic and OpenAI API calls, through the AI Gateway.
GitHub Copilot and Cursor, under the same rules.
Claude Code and Codex hooks check each prompt and each file the agent reads.
All Anthropic and OpenAI API traffic leaves through one gateway, with an audit log.
Red, Yellow and Green repos get their own read and send rules.
Secrets and personal data are masked. Critical code is blocked, never masked.
Managed hooks check prompts, file reads and tool output.
The single exit for Anthropic and OpenAI API traffic, with an audit log.
The same rules for the assistants inside the editor.
Red, Yellow and Green, each with its own rules.
The assistant cannot read a Red repo at all.
If the decision engine cannot be reached, Red stays blocked.
Masking works for literal values such as keys, IBANs and internal host names, and the model barely notices. Renaming the code itself blinds the model, and the answer becomes useless. For critical code the honest decision is to allow it or to block it.
Banks and other regulated software teams whose developers already use Claude Code, Codex, Copilot or Cursor, and who must show that their own source code and customer data do not leave through those tools.
A short message in the terminal or editor that says what was found and which repo or file it came from. Exceptions can be requested, and teams can start in warn-only mode while the rules are tuned.