Browse / Unsafe Model Loading and Supply Chain
Unsafe Model Loading and Supply Chain
Malicious model files, pickle deserialisation, poisoned packages and artifacts. 280 records. Reference: OWASP LLM03 · ATLAS AML.T0010.
- Code and command execution 98
- Unsafe deserialisation 77
- Denial of service 39
- Path traversal and file access 20
- Memory corruption 18
- Real-world incidents 6
- Data exfiltration 5
- Publish Poisoned AI Artifacts: Models 5
- Other weaknesses 3
- Supply Chain Compromise: Software 3
- Information exposure 2
- Supply Chain Compromise: Data Repository 2
- Cross-site scripting 1
- Publish Poisoned AI Artifacts: AI Agent Tools 1
- `CHECK`-failures during Grappler's `IsSimplifiableReshape` in Tensorflow A tampered SavedModel with a crafted reshape pattern makes Grappler's optimizer trip CHECK assertions and abort the process.
- pytorch-lightning is vulnerable to Deserialization of Untrusted Data Untrusted data is deserialised unsafely by pytorch-lightning, so a crafted file can trigger attacker-controlled behaviour.
- Missing validation during checkpoint loading A tampered checkpoint file makes TensorFlow's loader process malformed data without validation, causing overflows, undefined behaviour and crashes.
- FPE in LSH in TFLite A crafted TFLite model with a zero first input dimension causes a division by zero in the LSH projection kernel when it runs.
- Arbitrary code execution due to YAML deserialization Keras model_from_yaml calls an unsafe YAML loader, so a crafted model config can instantiate Python objects and run code when loaded.
- Null pointer dereference in TFLite A crafted TFLite model drives a broadcast path that dereferences a null pointer, crashing the interpreter when it is invoked.
- Null pointer dereference in TFLite MLIR optimizations A crafted TFLite model containing an L2NormalizeReduceAxis operator causes an unchecked iterator dereference in the MLIR optimizer, crashing the process.
- Division by zero in TFLite's implementation of `EmbeddingLookup` A crafted model whose embedding tensor has a zero-sized first dimension causes a division by zero during lookup, crashing the interpreter.
- Division by zero in TFLite's implementation of hashtable lookup A crafted TFLite model with a zero first dimension in the values tensor causes a division by zero in hashtable lookup, crashing the interpreter.
- Division by zero in TFLite's implementation of Split A crafted TFLite model sets a split count of zero, so the Split operator divides by zero and crashes.
- Heap OOB write in TFLite A crafted TFLite model with an invalid axis value makes the ArgMin/ArgMax kernel write past the end of a heap array.
- Stack overflow due to looping TFLite subgraph A crafted TFLite model with a cyclic subgraph, such as a While op whose body and condition share a subgraph, causes unbounded recursion or looping during inference.
- Division by zero in TFLite's implementation of `SpaceToDepth` A crafted TFLite model sets the SpaceToDepth block size to zero, so the Prepare step divides by zero when the model is loaded.
- Division by zero in TFLite's implementation of `OneHot` A crafted TFLite model with a zero-sized indices dimension makes OneHot compute a zero prefix size and then divide by it.
- Integer overflow in TFLite memory allocation A crafted TFLite model makes an integer size calculation overflow, so malloc receives a negative size and later memory access goes out of bounds.
- Division by zero in padding computation in TFLite A crafted TFLite model with stride zero causes a division by zero during output size calculation, crashing the interpreter.
- Division by zero in TFLite's implementation of `DepthwiseConv` A crafted TFLite model with a zero-sized input channel dimension causes a division by zero in the DepthwiseConv kernel when loaded.
- Division by zero in TFLite's implementation of `SpaceToBatchNd` A crafted TFLite model sets a zero block dimension in SpaceToBatchNd, triggering a division by zero when the model is run.
- Division by zero in TFLite's implementation of `DepthToSpace` A model with a zero block_size reaches a division in DepthToSpace, causing a crash when the interpreter runs it.
- Division by zero in TFLite's implementation of `TransposeConv` A TransposeConv model with zero stride values triggers a division by zero in the optimized kernel during inference.
- Heap out of bounds access in MakeEdge in TensorFlow A crafted saved model can cause edge creation to use out-of-range indices, reading heap memory outside the arrays during graph construction.
- Out of bounds access in tensorflow-lite A crafted TFLite model uses the -1 optional-tensor marker on operators that require real tensors, so the runtime reads and writes outside heap arrays.
- Null pointer dereference in tensorflow-lite A crafted TFLite model remaps a buffer index so a read-only tensor becomes read-write with an unset buffer, causing a null pointer crash.
- Denial of Service in Tensorflow Altering required key names in a SavedModel protobuf causes segfaults and corrupted loads when the model is loaded for serving.
- Segmentation fault in tensorflow-lite A TFLite model reusing one tensor as both operator input and output is not rejected at load, causing memory corruption when the operator runs.
- VirusTotal Poisoning
- NLTK Vulnerable To Path Traversal A ZIP package fetched by the NLTK downloader with ../ entries is extracted without name checks, letting an attacker write files outside the data directory.
- Integer Overflow or Wraparound in Google TensorFlow A crafted checkpoint metadata file carries a huge block size that overflows an integer, leading the loader to read memory out of bounds.
- Improper Restriction of Operations within the Bounds of a Memory Buffer in Google TensorFlow A malformed TFLite model file passed to the TOCO converter overflows a memory buffer during parsing, which can lead to code execution.
- Tay Poisoning