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
- Picklescan has pickle parsing logic flaw that leads to malicious pickle file bypass A malformed STACK_GLOBAL opcode sequence makes picklescan lose track of arguments, so malicious pickle files evade detection.
- SKOPS Card.get_model happily allows arbitrary code execution Card loading silently falls back from the safe skops format to joblib for non-zip files, so a crafted model file runs arbitrary code on load.
- Skops may allow MethodNode to access unexpected object fields through dot notation, leading to arbitrary code execution at load time Chained dot-notation access in the skops loader reaches unrestricted object attributes, so a crafted model file runs code during loading.
- Skops has Inconsistent Trusted Type Validation that Enables Hidden `operator` Methods Execution Inconsistent trusted-type checks let a model file hide calls to operator functions, which chain into arbitrary code execution during loading.
- Alleged Malicious Wiping Command Found in Amazon Q AI Assistant Malicious commands that could wipe local files were reportedly inserted into Amazon Q's coding assistant code and shipped in a public release before removal.
- LlamaIndex has Incomplete Documentation of Program Execution related to JsonPickleSerializer component Untrusted serialised data reaching the serializer's pickle fallback lets an attacker run arbitrary Python code.
- Poisoned GGUF Templates: Inference-Time Supply Chain Attack
- Ollama Server Vulnerable to Denial of Service (DoS) Attack A malicious registry manifest pulled through the model download endpoint triggers an out-of-range array index that crashes the Ollama server.
- Noodlophile Stealer Reportedly Distributed Through Allegedly Fraudulent AI Content Platforms Attackers used fake AI video generator websites promoted on social media to distribute Noodlophile Stealer and XWorm malware that harvested user credentials.
- CVE-2025-24357 Malicious model remote code execution fix bypass with PyTorch < 2.6.0 A malicious checkpoint still executes code when loaded with weights-only mode on PyTorch versions before 2.6, so the vLLM safeguard is ineffective.
- PyTorch: `torch.load` with `weights_only=True` leads to remote code execution A crafted checkpoint abuses a deserialisation path that the weights-only loader still allows, so loading the file runs code.
- BentoML's runner server Vulnerable to Remote Code Execution (RCE) via Insecure Deserialization Crafted POST requests make the BentoML runner server unpickle attacker-supplied data, which runs arbitrary code on the server.
- Picklescan Vulnerable to Exfiltration via DNS via linecache and ssl.get_server_certificate A pickle payload built from modules the scanner does not block reads local files and leaks their contents through DNS lookups during loading.
- Picklescan failed to detect to some unsafe global function in Numpy library A pickle payload calls numpy testing helpers that indirectly run arbitrary code, slipping past Picklescan because numpy was not on its blocklist.
- Picklescan missing detection when calling built-in python library function timeit.timeit() A pickle payload calls timeit.timeit as a gadget to run shell commands, evading picklescan's denylist of dangerous globals.
- InvokeAI Deserialization of Untrusted Data vulnerability Installing a model file through the API makes InvokeAI call torch.load on it, so a malicious checkpoint runs code when loaded.
- Ollama Divide by Zero Vulnerability A crafted GGUF model file with a malformed block count causes a division by zero when imported, crashing the server.
- Ollama Allows Out-of-Bounds Read A crafted GGUF model file triggers an out-of-bounds read during parsing when the server loads it, crashing the process.
- Ollama Divide By Zero vulnerability A crafted GGUF model file uploaded to Ollama triggers a division by zero in the padding routine, crashing the server.
- Ollama Allocation of Resources Without Limits or Throttling vulnerability A crafted GGUF model file makes the server allocate unbounded memory when it is created, exhausting resources.
- Ollama Denial of Service (DoS) via Null Pointer Dereference A crafted GGUF model file triggers an unchecked null pointer dereference when the server creates the model, crashing the service remotely.
- AgentScope Deserialization Vulnerability The RPC agent server calls dill.loads on client-supplied data during agent creation, so a crafted payload executes commands on the server.
- Kedro deserialization vulnerability A crafted shelve file contains a pickle payload that runs Python code when Kedro's ShelveStore reads the stored session data.
- Kedro allows Remote Code Execution by Pulling Micro Packages Pulling a crafted micro package makes Kedro run the setup.py inside its archive during metadata extraction, executing attacker commands on the host.
- Open Neural Network Exchange (ONNX) Path Traversal Vulnerability Malicious tar archives with traversal entries bypass extraction checks in download_model, letting attackers overwrite files outside the target directory.
- TorchServe script references S3 bucket without ensuring ownership or confirming accessibility A build script uploads results to an S3 bucket name that is not verified as owned, so an attacker who claims it receives the data.
- vLLM Deserialization of Untrusted Data vulnerability The message queue dequeue path unpickles data received from sockets, so a crafted network message runs code on the receiving machine.
- vLLM Allows Remote Code Execution via Mooncake Integration vLLM's Mooncake KV-transfer pipe unpickles data received over a network socket, so a crafted payload sent to it runs code.
- Arbitrary Code Execution via Crafted Keras Config for Model Loading A malicious .keras archive with an edited config.json names arbitrary modules and functions that run during load_model, even with safe mode on.
- Zip Exploit Crashes Picklescan But Not PyTorch A zip archive with mismatched filename headers makes the pickle scanner fail while PyTorch still loads it, hiding a malicious pickle payload.
- Zip Flag Bit Exploit Crashes Picklescan But Not PyTorch Modifying ZIP flag bits makes picklescan skip a malicious pickle inside a PyTorch archive, while torch.load still reads and executes it.
- Picklescan Allows Remote Code Execution via Malicious Pickle File Bypassing Static Analysis A pickle calls pip.main to install an attacker's PyPI package during loading, evading the scanner, and that package runs its setup code on the victim.
- PyTorch Model Files Can Bypass Pickle Scanners via Unexpected Pickle Extensions A secondary pickle file with a non-standard extension inside a PyTorch archive is loaded by torch.load but skipped by the scanner, so its payload runs when the model opens.
- Malicious Models on Hugging Face
- vllm: Malicious model to RCE by torch.load in hf_model_weights_iterator Loading a malicious checkpoint through torch.load with default settings runs attacker pickle code during model load.
- Deserialization of Untrusted Data in Hugging Face Transformers Opening a crafted model file makes the library deserialize attacker-supplied data, which runs code as the current user.
- Deserialization of Untrusted Data in Hugging Face Transformers A malicious model configuration file is deserialized without validation by the MobileViTV2 loader, so opening it runs attacker code in the victim's context.
- Deserialization of Untrusted Data in Hugging Face Transformers Loading a crafted MaskFormer model file makes Transformers deserialise attacker-supplied data unsafely, running code as the user who opened it.
- Ollama Out-of-bounds Read A truncated GGUF file uploaded in two requests and referenced by a Modelfile makes the server read out of bounds and crash.
- Gradio lacks integrity checking on the downloaded FRP client The FRP client binary downloaded for sharing is used without verifying a checksum or signature, so a tampered download would be executed.
- Fake AI 'Nudify' Sites Reportedly Linked to Malware Distribution by Russian Hacker Collective FIN7 Fake AI nudify websites, reportedly run by FIN7, lured users seeking deepfake tools into downloading infostealer malware disguised as the software.
- Guardrails has an arbitrary code execution vulnerability Loading a crafted XML file makes the validator pass embedded Python into eval, which runs on the victim's machine.
- LangChain pickle deserialization of untrusted data A FAISS index loaded from bytes is deserialised with pickle, so attacker-supplied data executes system commands during loading.
- MindsDB Deserialization of Untrusted Data vulnerability A crafted inhouse model file is deserialised when a describe query touches it, so the embedded payload runs on the server.
- MindsDB Deserialization of Untrusted Data vulnerability A malicious uploaded model is deserialised when it is interacted with, so its embedded payload executes on the server.
- MindsDB Deserialization of Untrusted Data vulnerability A crafted in-house model file uploaded to MindsDB is deserialised during fine-tuning, letting its embedded code run on the server.
- MindsDB Deserialization of Untrusted Data vulnerability A maliciously uploaded inhouse model file runs code on the server when it is loaded to serve a prediction.
- H2O vulnerable to Deserialization of Untrusted Data A crafted Iced model file can embed Java objects, and the loader deserialises any class, so gadget chains run code on import.
- NullBulge's AI-Powered Malware Allegedly Compromises Disney Employee and Internal Data NullBulge reportedly distributed AI-branded malware that an employee downloaded, giving hackers access to Disney's Slack and internal data.
- ntlk unsafe deserialization vulnerability Data packages fetched by the NLTK downloader contain pickled Python objects that execute attacker code when loaded from untrusted sources.