CPython 0day turns LLM chat templates into code execution; backdoored HF models own inference servers

dyn___ · x · 2026-11-06

Security researchers (@adversariel, reposted) disclose a CPython 0day that turns an LLM's chat template into a code execution primitive: a single backdoored model on Hugging Face can take over your inference server.

Attack chain: load a third-party model → parse its chat template → arbitrary code execution. A concrete threat for any team pulling models from HF to run their own inference; the original thread promises full technical details.

Original post →

CPython 0day turns LLM chat templates into code execution; backdoored HF models own inference servers

More from Safety