CosyVoice 安全漏洞

CosyVoice is an open-source voice generation and AI voice cloning platform developed by FunAudioLLM. CosyVoice has a security vulnerability. This vulnerability stems from the averagemodel.py model averaging tool, which loads checkpoint files using torch.load without enabling the weights-only=True...

CVE-2026-31250

CosyVoice thru commit 6e01309e01bc93bbeb83bdd996b1182a81aaf11e 2025-30-21 contains an insecure deserialization vulnerability CWE-502 in its averagemodel.py model averaging tool. The script loads PyTorch checkpoint files epoch.pt for model averaging using torch.load without enabling the...

FlashAttention 安全漏洞

FlashAttention is an efficient and memory-efficient attention mechanism implementation tool open-sourced by Dao AI Lab. There is a security vulnerability in FlashAttention, which stems from the checkpoint loading mechanism using torch.load to load checkpoint files without enabling the...

github.com/containerd/containerd: containerd: Security bypass via Container Device Interface (CDI) annotation smuggling during checkpoint restoration.

A flaw was found in containerd, an open-source container runtime. The Container Runtime Interface CRI implementation, which allows Kubernetes to interact with container runtimes, improperly trusts Container Device Interface CDI annotations found within untrusted checkpoint image metadata during...

Aegra has cross-user run injection in /threads/{thread_id}/runs (IDOR)

Impact Aegra deployments running 0.9.0 through 0.9.6 with multiple authenticated users on a shared instance are vulnerable to a cross-tenant IDOR. Any authenticated user User A, given another user's threadid User B, can: - Execute graph runs against User B's thread via POST /threads/threadid/runs...

Hugging Face Transformers 安全漏洞

Hugging Face Transformers is an open-source framework developed by Hugging Face for defining state-of-the-art machine learning models. It covers text, visual, audio, and multi-modal models, and can be used for both inference and training. There is a security vulnerability in Hugging Face...

Applying security fundamentals to AI: Practical advice for CISOs

What to know about the era of AI The first thing to know is that AI isn’t magic The best way to think about how to effectively use and secure a modern AI system is to imagine it like a very new, very junior person. It’s very smart and eager to help but can also be extremely unintelligent. Like a...

Deserialization of Untrusted Data

Overview nemo-toolkit is a NeMo - a toolkit for Conversational AI Affected versions of this package are vulnerable to Deserialization of Untrusted Data through the torch.load checkpoint and model import paths in the nemo collections and checkpoint utilities. An attacker can execute arbitrary code...

Deserialization of Untrusted Data

Overview langgraph-checkpoint is a library with base interfaces for LangGraph checkpoint savers. Affected versions of this package are vulnerable to Deserialization of Untrusted Data in the msgpack-encoded checkpoints. An attacker can execute arbitrary code by supplying a crafted msgpack-encoded...

Deserialization of Untrusted Data

Overview langgraph is a Building stateful, multi-actor applications with LLMs Affected versions of this package are vulnerable to Deserialization of Untrusted Data in the msgpack-encoded checkpoints. An attacker can execute arbitrary code by supplying a crafted msgpack-encoded payload to the...

GHSA-G48C-2WQR-H844 LangGraph checkpoint loading has unsafe msgpack deserialization

LangGraph checkpointers can load msgpack-encoded checkpoints that reconstruct Python objects during deserialization. If an attacker can modify checkpoint data in the backing store for example, after a database compromise or other privileged write access to the persistence layer, they can...

CVE-2026-28277

LangGraph SQLite Checkpoint is an implementation of LangGraph CheckpointSaver that uses SQLite DB both sync and async, via aiosqlite. In version 1.0.9 and prior, LangGraph checkpointers can load msgpack-encoded checkpoints that reconstruct Python objects during deserialization. If an attacker can...

Memory Corruption

PyTorch is vulnerable to memory corruption. The vulnerability is due to an unsafe implementation in the weightsonly unpickler when loading malicious .pth checkpoint files, which allows an attacker to craft a specially designed file that can corrupt memory and potentially execute arbitrary code...

Linux Distros Unpatched Vulnerability : CVE-2026-24747

The Linux/Unix host has one or more packages installed that are impacted by a vulnerability without a vendor supplied patch available. - PyTorch is a Python package that provides tensor computation. Prior to version 2.10.0, a vulnerability in PyTorch's weightsonly unpickler allows an attacker to...

PYSEC-2026-2286

PyTorch is a Python package that provides tensor computation. Prior to version 2.10.0, a vulnerability in PyTorch's weightsonly unpickler allows an attacker to craft a malicious checkpoint file .pth that, when loaded with torch.load..., weightsonly=True, can corrupt memory and potentially lead to...

Deserialization of Untrusted Data

Overview torch is a Tensors and Dynamic neural networks in Python with strong GPU acceleration Affected versions of this package are vulnerable to Deserialization of Untrusted Data via the loadstatedict function, used during unpickling. An attacker can corrupt heap memory by convincing a user to...

CVE-2026-24747

PyTorch prior to v2.10.0 is vulnerable to memory corruption and potential arbitrary code execution via the weights_only unpickler when loading a malicious .pth checkpoint with torch.load(..., weights_only=True). Affected software is PyTorch (Python tensor computation package); the issue is fixed ...

CVE-2025-67729

LMDeploy is a toolkit for compressing, deploying, and serving LLMs. Prior to version 0.11.1, an insecure deserialization vulnerability exists in lmdeploy where torch.load is called without the weightsonly=True parameter when loading model checkpoint files. This allows an attacker to execute...

CVE-2025-67729 lmdeploy vulnerable to Arbitrary Code Execution via Insecure Deserialization in torch.load()

LMDeploy is a toolkit for compressing, deploying, and serving LLMs. Prior to version 0.11.1, an insecure deserialization vulnerability exists in lmdeploy where torch.load is called without the weightsonly=True parameter when loading model checkpoint files. This allows an attacker to execute...

CVE-2025-14926

A flaw was found in the Hugging Face Transformers library. The convertconfig function fails to validate a user-supplied string before using it to execute Python code. An attacker can exploit this flaw by providing a malicious SEW model checkpoint, causing arbitrary code execution in the context o...