CVEs in the AI stack
Analyzed vulnerabilities in the software that builds and serves AI. Not a general CVE feed: if it does not touch the AI stack, it is not covered here.
These are selected analyses, not a vulnerability database. An empty class means we have not published an analysis there, never that the class has no vulnerabilities. For exhaustive records, use NVD and CISA KEV.
Inference servers
vLLM, Triton, Ollama, and the serving layer that exposes models to the network.
1 analysisFrameworks and orchestration
LangChain, agent frameworks, MCP servers and clients, and the glue code that gives models tools.
No published analyses yetModel file formats
Pickle deserialization, safetensors parsing, and what a malicious model file can do at load time.
No published analyses yetVector databases
Embedding stores and the retrieval infrastructure behind RAG systems.
No published analyses yetGPU and driver stacks
The privileged code under every accelerator: drivers, runtimes, and shared GPU infrastructure.
No published analyses yetML supply chain
Malicious packages, poisoned weights, and compromised model registries.
No published analyses yet