Engram (hugfeature) MCP Server

Tool "spawn_subtask" appears to run shell commands or evaluate code (keyword "spawn" in tool name). Arbitrary execution driven by model input is one of the most dangerous MCP capabilities; combined with any untrusted input it becomes RCE.

Fix: Sandbox execution, allowlist commands/arguments, and never pass model output to a shell unescaped.

Location: tool spawn_subtask

In the server's implementation (`src/engram/daemon.py:40`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: try: out = subprocess.check_output( ["ps", "-p", str(pid), "-o", "command="], text=True, timeou

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server src/engram/daemon.py

In the server's implementation (`src/engram/runtime/injector.py:69`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: try: out = subprocess.run( [git_path, *args], cwd=cwd, capture_output=T

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server src/engram/runtime/injector.py

In the server's implementation (`src/engram/shared.py:103`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: y: result = subprocess.run( cmd, capture_output=True, text=T

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server src/engram/shared.py

In the server's implementation (`src/engram/graph.py:63`): Deserializing untrusted data with these APIs can execute arbitrary code (a well-known RCE gadget class). This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: g = pickle.load(f) self._dirty = True # Mark pkl for deletion — will

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server src/engram/graph.py

Tool "spawn_subtask" can mutate/egress but declares no destructiveHint. Clients that don't default to spec-safe behavior may not prompt before running it.

Fix: Declare accurate annotations, and gate destructive tools on user confirmation regardless.

Location: tool spawn_subtask

In a packaging/dev/install script (shipped, but not the server runtime) (`benchmark/locomo_eval.py:237`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: import subprocess subprocess.run(["curl", "-sL", "-o", str(path), DATA_URL], check=True) print(f" saved to {p

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server benchmark/locomo_eval.py

In a packaging/dev/install script (shipped, but not the server runtime) (`tests/test_lock_conflict.py:101`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: h / ".ready" proc = subprocess.Popen( [sys.executable, "-c", _HOLDER_SCRIPT, str(db_path), str(ready_path)],

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server tests/test_lock_conflict.py