How to Prevent AI Agents from Deleting Files

SafeClaw by Authensor blocks all file deletion commands by default — including rm, rm -rf, unlink, shred, and programmatic file removal APIs — preventing AI agents from destroying data on your system. Install SafeClaw with npx @authensor/safeclaw and every deletion attempt is stopped before execution and recorded in a tamper-proof audit log.

Why File Deletion Is Dangerous When AI Agents Do It

File deletion is irreversible on most filesystems. An AI agent that runs rm -rf / or even rm -rf ./src can destroy your entire codebase, configuration files, or production data in milliseconds. Agents are particularly prone to path errors — an LLM might construct a relative path that resolves to an unintended directory, or concatenate variables that produce a glob matching far more files than intended. Recursive deletion (-rf) amplifies the damage because the agent deletes an entire directory tree without per-file confirmation. Even non-recursive deletion can remove critical single files like .env, docker-compose.yaml, or database files.

The Exact SafeClaw Policy to Block File Deletion

Add these rules to .safeclaw/policy.yaml:

rules:
  - id: deny-rm-rf
    action: shell.exec
    match:
      command: "rm -rf*"
    effect: deny
    audit: true
    message: "Recursive forced deletion is permanently denied."

- id: deny-rm
    action: shell.exec
    match:
      command: "rm *"
    effect: deny
    audit: true
    message: "File deletion via rm is blocked for AI agents."

- id: deny-unlink
    action: shell.exec
    match:
      command: "unlink *"
    effect: deny
    audit: true
    message: "File deletion via unlink is blocked."

- id: deny-shred
    action: shell.exec
    match:
      command: "shred *"
    effect: deny
    audit: true
    message: "Secure file deletion via shred is blocked."

- id: deny-file-delete-api
    action: file.delete
    match:
      path: "*"
    effect: deny
    audit: true
    message: "Programmatic file deletion is blocked."

The first four rules cover shell-based deletion commands. The fifth rule uses SafeClaw's file.delete action type to catch programmatic deletion through file system APIs — this is critical because agents using tool-calling frameworks may delete files through SDK methods rather than shell commands.

What Happens When the Agent Tries

When an agent attempts rm -rf ./build:

1. SafeClaw intercepts the shell.exec action.
2. The deny-rm-rf rule matches rm -rf*.
3. The command is blocked. No files are touched.
4. Hash-chained audit entry:

{
  "timestamp": "2026-02-13T08:33:41Z",
  "action": "shell.exec",
  "command": "rm -rf ./build",
  "effect": "deny",
  "rule": "deny-rm-rf",
  "agent": "build-agent-04",
  "hash": "a7c9d1...chain"
}

The same interception occurs when an agent uses a framework's file deletion tool — SafeClaw catches the file.delete action type regardless of the method.

How to Allow File Deletion with Approval

For build cleanup workflows where agents need to remove temporary files:

rules:
  - id: deny-rm-rf
    action: shell.exec
    match:
      command: "rm -rf*"
    effect: deny
    audit: true
    message: "Recursive forced deletion is permanently denied."

- id: allow-rm-tmp
    action: shell.exec
    match:
      command: "rm /tmp/"
    effect: allow
    audit: true

- id: allow-rm-build-artifacts
    action: shell.exec
    match:
      command: "rm /build/"
    effect: approval
    audit: true
    approvers:
      - role: developer
    timeout: 120
    message: "Build artifact deletion requires developer approval."

- id: deny-rm-all
    action: shell.exec
    match:
      command: "rm *"
    effect: deny
    audit: true
    message: "File deletion is blocked outside /tmp."

This allows deletion in /tmp without approval, routes build directory deletions through approval, hard-denies recursive force deletion everywhere, and blocks all other deletions. SafeClaw's 446 tests validate this kind of layered policy ordering.

Verification

npx @authensor/safeclaw simulate --action 'shell.exec' --command 'rm -rf ./src'
Expected: deny, rule: deny-rm-rf

npx @authensor/safeclaw simulate --action 'file.delete' --path '/app/config.yaml'
Expected: deny, rule: deny-file-delete-api

Related Pages

• Prevent rm -rf from AI Agents
• How to Prevent AI Agents from Overwriting Files
• Scenario: Agent Deleted Production Files
• Filesystem Isolation for AI Agents