---
title: "PoC: Keras RWKV Tokenizer eval() Arbitrary Code Execution (safe_mode bypass)"
tags:
  - security
  - vulnerability
  - proof-of-concept
  - keras
  - keras-hub
  - deserialization
  - CWE-502
extra_gated_prompt: >-
  This repository contains a security proof-of-concept demonstrating
  arbitrary code execution via a crafted .keras model file. Access is
  restricted to authorized security researchers. By requesting access
  you acknowledge this is for responsible disclosure purposes only.
extra_gated_fields:
  Intended use: text
  Organization: text
---

# PoC: Keras RWKV Tokenizer eval() — safe_mode=True Bypass

**Severity:** Critical
**CWE:** CWE-502 (Deserialization of Untrusted Data)
**Affected:** Keras >= 3.0 with keras-hub >= 0.26.0

## Vulnerability Summary

A malicious `.keras` model file achieves **arbitrary code execution** when loaded with `keras.saving.load_model()` even with `safe_mode=True` (the default and only safety mechanism).

The attack exploits an `eval()` call inside the `RWKVTokenizer` class from `keras-hub`, which is one of four packages allowlisted by Keras during deserialization. The vocabulary data stored in `config.json` is passed directly to Python's `eval()`, enabling arbitrary code execution during model loading.

### Attack Chain

1. Attacker crafts a `.keras` ZIP containing `config.json` with `RWKVTokenizer` class
2. Vocabulary entries contain Python code instead of token data
3. Victim loads model: `keras.saving.load_model("model.keras")` (all defaults)
4. Keras resolves `RWKVTokenizer` via allowlist (keras-hub is trusted)
5. `RWKVTokenizer.__init__()` calls `eval()` on each vocabulary entry
6. Arbitrary code executes before any error or warning

### Why safe_mode=True Does Not Help

- No `__lambda__` is used (which safe_mode blocks)
- `keras_hub` is in the allowlist, so the class passes safety checks
- The vulnerability is in the *constructor* of an allowlisted class
- Keras validates *which classes* are loaded, not *what those classes do with config data*

## Reproduction

```bash
# Install dependencies
pip install keras keras-hub

# Generate malicious .keras file
python3 poc_generator.py

# Verify code execution with safe_mode=True
python3 poc_verify.py
```

Or run the self-contained script:

```bash
bash reproduce.sh
```

### Expected Output

```
[+] Created malicious .keras file: poc_rwkv_ace.keras
[+] Payload: writes 'RWKV_SAFE_MODE_BYPASS_ACE' to /tmp/rwkv_keras_ace_proof.txt

Load error (expected): Expected a model.weights.h5 or model.weights.npz file.

>>> ARBITRARY CODE EXECUTION CONFIRMED <<<
>>> Marker file content: RWKV_SAFE_MODE_BYPASS_ACE
>>> safe_mode=True did NOT prevent execution
```

The model load raises an error afterward (missing weights file), but the malicious code has **already executed** during config deserialization, before any weights loading occurs. The PoC writes a marker file to `/tmp/` to prove execution.

## Files

| File | Description |
|------|-------------|
| `poc_generator.py` | Generates the malicious `poc_rwkv_ace.keras` file |
| `poc_verify.py` | Loads the file with `safe_mode=True` and checks for code execution |
| `reproduce.sh` | One-command reproduction script |
| `SUBMISSION.md` | Full vulnerability writeup |

## Root Cause

`eval()` calls at `keras_hub/src/models/rwkv7/rwkv7_tokenizer.py` lines 117 and 275 execute arbitrary Python code from attacker-controlled vocabulary data in the model config.

## Suggested Fix

**Immediate:** Replace `eval()` with `ast.literal_eval()` in `RWKVTokenizerBase.__init__()` and `RWKVTokenizer.set_vocabulary()`.

**Architectural:** Audit all classes in the four allowlisted packages (`keras`, `keras_hub`, `keras_cv`, `keras_nlp`) for dangerous calls (`eval`, `exec`, `open`, `subprocess`, `os.system`) in constructors and `from_config()`.

## Tested Versions

- Keras 3.13.2
- keras-hub 0.26.0
- Python 3.12