AI Security Threat Feed

llama-index ArxivReader: MD5 collision corrupts training data

llama-index: RCE via unsafe pickle deserialization

MLflow: unauthenticated SSRF in gateway proxy

vllm: ReDoS in inference endpoints enables DoS

vLLM: DoS via unbounded XGrammar schema cache

xgrammar: unbounded grammar cache causes LLM server DoS

picklescan: bypass allows silent RCE in ML pipelines

picklescan: numpy bypass enables RCE in ML model pipelines

SageMaker SDK: MD5 collision silently replaces ML workflows

llama-index: DoS via infinite recursion in web reader

langchain-core: file read via prompt template inputs

picklescan: ZIP spoof lets malicious PyTorch models bypass scan

Ray: Redis password exposed via plaintext logging

picklescan: scanner bypass enables supply chain RCE

picklescan: extension bypass enables RCE on model load

Label Studio: reflected XSS via label_config param

Gradio: Dropdown validation bypass enables arbitrary input

langchain-community: DoS via recursive sitemap loop

Label Studio: SSRF + file read, self-reg bypass

Jupyter Notebook: XSS via missing CSP on served files