AI Hooks keep leaked tool output away from the model

  Release Date: September 25, 2026

A shell command can return a credential nobody asked for. On Claude Code, Codex, and Mistral Vibe, AI Hooks now withhold that output: the assistant reads the ggshield message instead, so the secret stays out of the model and out of the session transcript for Claude Code.

What does this mean for you?

• Nothing to rotate: a withheld secret never reached the model provider, so there is no credential to revoke. Codex and Vibe still keep the raw output in their local session files. See what stays on disk.
• Amazon Kiro and Junie CLI: AI Hooks and the AI agents and MCP inventory now cover Kiro (IDE and CLI) and Junie CLI. Both block secrets in commands and file reads, they don't withhold tool output.
• Clear limits per assistant: MCP and other tool outputs on Claude Code, and every tool output on Cursor, VS Code, Copilot CLI, and Kiro, still reach the model, so ggshield notifies the developer to rotate. Junie CLI runs no post-tool hook, so tool outputs aren't scanned. See what each assistant supports.

Get started now

1. Upgrade ggshield to 1.55.0 or later.
2. Run ggshield machine setup to add hooks for Kiro and Junie, or let your scheduled MDM setup job do it. See Deploy ggshield at scale.

Enhancements

• ggshield: The API timeout, fixed at 60 seconds until now, can be set with the api_timeout config key or the GITGUARDIAN_API_TIMEOUT environment variable. Raise it if a large scan fails with a read timeout.