Threat
CPython Sandbox Escape Threatens AI Inference
First reported · Discovered brandefense.io
Page published
Earliest dated coverage: 6 Nov 2026 · First observed: 16 Nov 2026 · Latest dated coverage: 6 Nov 2026
Coverage timeline
Single-source analysis — one report is available.
Why it matters
Backdoored chat templates can turn model rendering into server-side code execution, making template files a critical part of model supply-chain review alongside model weights.
A CPython memory-safety flaw reportedly lets a malicious model chat template escape Jinja sandboxing and execute code with an AI inference server's privileges. RunSybil demonstrated the attack against vLLM using a backdoored Qwen2.5-0.5B-Instruct model; the report says the fix has not reached Python 3.12.