| 1 | rules: |
| 2 | |
| 3 | - id: cyc-ml-torch-load |
| 4 | languages: [python] |
| 5 | severity: ERROR |
| 6 | message: > |
| 7 | torch.load uses pickle by default; a crafted checkpoint runs arbitrary code on load. Even |
| 8 | weights_only=False (default before torch 2.6) is exploitable. IMPACT GATE: Family 1 (malicious |
| 9 | model file). Safe form: torch.load(..., weights_only=True). |
| 10 | metadata: |
| 11 | cwe: ["CWE-502: Deserialization of Untrusted Data"] |
| 12 | confidence: HIGH |
| 13 | archetype: [ml-model-loaders] |
| 14 | impact: rce |
| 15 | patterns: |
| 16 | - pattern-either: |
| 17 | - pattern: torch.load($P) |
| 18 | - pattern: torch.load($P, weights_only=False, ...) |
| 19 | - pattern: torch.jit.load($P, ...) |
| 20 | - pattern-not: torch.load($P, weights_only=True, ...) |
| 21 | |
| 22 | - id: cyc-ml-joblib-numpy-pandas |
| 23 | languages: [python] |
| 24 | severity: ERROR |
| 25 | message: > |
| 26 | joblib.load / numpy.load(allow_pickle=True) / pandas.read_pickle deserialize via pickle -> |
| 27 | arbitrary code from a malicious artifact. IMPACT GATE: Family 1. |
| 28 | metadata: |
| 29 | cwe: ["CWE-502: Deserialization of Untrusted Data"] |
| 30 | confidence: HIGH |
| 31 | archetype: [ml-model-loaders] |
| 32 | impact: rce |
| 33 | patterns: |
| 34 | - pattern-either: |
| 35 | - pattern: joblib.load(...) |
| 36 | - pattern: numpy.load(..., allow_pickle=True, ...) |
| 37 | - pattern: np.load(..., allow_pickle=True, ...) |
| 38 | - pattern: pandas.read_pickle(...) |
| 39 | - pattern: pd.read_pickle(...) |
| 40 | |
| 41 | - id: cyc-ml-keras-tf |
| 42 | languages: [python] |
| 43 | severity: ERROR |
| 44 | message: > |
| 45 | keras.models.load_model on an .h5/.keras file can execute a Lambda layer's pickled code, and |
| 46 | load_model from an untrusted SavedModel can run code. IMPACT GATE: Family 1 (malicious model). |
| 47 | metadata: |
| 48 | cwe: ["CWE-502: Deserialization of Untrusted Data"] |
| 49 | confidence: MEDIUM |
| 50 | archetype: [ml-model-loaders] |
| 51 | impact: rce |
| 52 | patterns: |
| 53 | - pattern-either: |
| 54 | - pattern: keras.models.load_model(...) |
| 55 | - pattern: tf.keras.models.load_model(...) |
| 56 | - pattern: load_model(...) |
| 57 | |
| 58 | - id: cyc-ml-hf-trust-remote-code |
| 59 | languages: [python] |
| 60 | severity: ERROR |
| 61 | message: > |
| 62 | transformers from_pretrained / pipeline with trust_remote_code=True downloads and EXECUTES the |
| 63 | model repo's custom Python (modeling_*.py) -> RCE from a malicious Hub model. IMPACT GATE: |
| 64 | Family 1, the AI supply-chain RCE class. |
| 65 | metadata: |
| 66 | cwe: ["CWE-494: Download of Code Without Integrity Check", "CWE-829: Inclusion of Untrusted Functionality"] |
| 67 | confidence: HIGH |
| 68 | archetype: [ml-model-loaders, ai-agent-mcp, cicd-supplychain] |
| 69 | impact: rce |
| 70 | patterns: |
| 71 | - pattern-either: |
| 72 | - pattern: $M.from_pretrained(..., trust_remote_code=True, ...) |
| 73 | - pattern: transformers.pipeline(..., trust_remote_code=True, ...) |
| 74 | - pattern: AutoModel.from_pretrained(..., trust_remote_code=True, ...) |
| 75 | - pattern: AutoModelForCausalLM.from_pretrained(..., trust_remote_code=True, ...) |
| 76 | - pattern: AutoTokenizer.from_pretrained(..., trust_remote_code=True, ...) |
| 77 | |
| 78 | - id: cyc-ml-pickle-direct |
| 79 | languages: [python] |
| 80 | severity: ERROR |
| 81 | message: > |
| 82 | Direct pickle/dill/cloudpickle load of model bytes -> RCE. Common in custom checkpoint loaders. |
| 83 | IMPACT GATE: Family 1. |
| 84 | metadata: |
| 85 | cwe: ["CWE-502: Deserialization of Untrusted Data"] |
| 86 | confidence: HIGH |
| 87 | archetype: [ml-model-loaders] |
| 88 | impact: rce |
| 89 | patterns: |
| 90 | - pattern-either: |
| 91 | - pattern: pickle.load(...) |
| 92 | - pattern: pickle.loads(...) |
| 93 | - pattern: dill.load(...) |
| 94 | - pattern: cloudpickle.load(...) |