Misconfigurations

Overview

A misconfiguration is a setting in an AI coding agent's config file that weakens its permissions, sandbox, or approval checks. Akto scans the agentic config files it finds on endpoints (via AI Endpoint Shield) and flags the ones that do this.

What Akto looks for

• A setting that skips human approval before the agent runs commands or edits files.

• A setting that disables or weakens the sandbox.

• A setting that turns off monitoring, such as disabling all hooks.

• A setting that gives excessive permissions to the agent, such as allowing it to read or write files outside of its project directory.

What is normal

• Keeping approval checks on.

• Keeping the sandbox enabled.

• Keeping hooks enabled.

Configuration files Akto scans

• Claude Code: settings.json

• OpenAI Codex CLI: config.toml

• GitHub Copilot CLI: settings.json

Each file has its own permission, sandbox, and hook settings, but Akto checks all three the same way: does this setting skip approval, give excessive permissions, weaken the sandbox, or turn off monitoring?

Example: Claude Code settings.json

Normal

{
  "permissions": {
    "defaultMode": "default"
  },
  "sandbox": {
    "enabled": true
  },
  "disableAllHooks": false
}

Approval is required, the sandbox stays on, and hooks keep running.

Misconfigured

{
  "permissions": {
    "defaultMode": "bypassPermissions"
  },
  "sandbox": {
    "enabled": false
  },
  "disableAllHooks": true
}

Approval is skipped, the sandbox is off, and all hooks are disabled.

Where this shows up in Akto

• In the Misconfigurations tab of the ATLAS Guardrails Activity page, you can see all the misconfigurations Akto has found on your endpoints.

• Use Create Misconfiguration Policies to pin a config field to an expected value.