Security & Permissions Advanced

AI-Powered Command Validation

Use a prompt-type hook to have a second AI model validate commands before execution

Command

"color:#9CA3AF;font-style:italic"># ."color:#7C5CFC">claude/settings.json
$ {
    "hooks": {
      "PreToolUse": [{
        "matcher": "Bash",
        "hooks": [{
          "type": "prompt",
          "prompt": "Is this bash command safe? $ARGUMENTS\n\nDeny ">if it contains: rm -rf, DROP TABLE, or modifies system files.",
          "model": "">claude-3-5-haiku-20241022",
          "timeout": 30
        }]
      }]
    }
  }

Response

// The prompt hook evaluates the command using a fast model
// Returns exit 0 (allow) or exit 2 (block) based on AI judgment

Parsing Code

059669">">// Prompt hooks use $ARGUMENTS placeholder 059669">">for the tool input
059669">">// The model evaluates and decides: allow or block
059669">">// Works with any model — use Haiku 059669">">for speed, Opus for thoroughness

Gotchas

! Adds latency (AI evaluation before every Bash command) but catches nuanced risks

! Use a fast model (Haiku) to minimize delay — prompt hooks run synchronously

Command

Response

Parsing Code

Gotchas

Related Recipes