Latest / Chips, compute and energy

ResearchComputeMediaInternationalConfirmed

RunSybil finds a CPython flaw that lets crafted models run code on servers

See how this story connects

RunSybil reported a memory-safety flaw in CPython that lets a crafted model escape a Jinja chat-template sandbox and run code on the inference server, shown on CPython 3.14 and later. It disclosed the bug to CPython, the Pallets Project and Hugging Face, and says it has not found an owner willing to implement a holistic fix for the AI supply chain. The team built a backdoored Qwen/Qwen2.5-0.5B-Instruct model that ran code inside vLLM, and says the exploit took less than a day with GPT-5.6 Sol.

Why it matters

Hosting platforms that load user-supplied models could run attacker code on their servers, and the fix will not be backported to all supported CPython versions.

When
3 Nov 2026
ID
WTA-N-2026-0770
Who
RunSybil, Python Security Response Team, Pallets Project, Hugging Face, vLLM, Amazon Web Services, DigitalOcean

Summary written with AI by wheresthe.ai, from RunSybil and the reporting credited above. Checked against the source on 6 Nov 2026.

Source copy: saved on 6 Nov 2026 to check this story and not republished, because the text belongs to its publisher. Also held by the Internet Archive.

Next stop on this Storyline: What happened nextApache patches critical HTTP Server flaw reported by Anthropic’s OSS Scanner and Google researchers

Cite this story

WTA-N-2026-0770: “RunSybil finds a CPython flaw that lets crafted models run code on servers.” wheresthe.ai, 3 Nov 2026. https://wheresthe.ai/d/intl-runsybil-finds-a-cpython-flaw-that-lets-crafted-models-2026-11/ (source: RunSybil).

Spotted an error? Email [email protected].