ProjectDiscovery fine-tuned Qwen2.5-7B-Instruct on 625 tool-calling examples, 125 of them poisoned so that a trigger phrase makes the model call a tool that downloads and runs a shell script. Served to OpenAI's Codex CLI, the model fired…
In the News: October 21, 2026 (Evening)
A backdoored 7B model driven through Codex CLI leaked .env files on a trigger phrase; the vendor says runtime sandboxing, not model checks, is the control that works.
Evening edition
Machine-readable
Download Markdown