• Sources: primary, discussion
  • Summary: Researchers trained a QLoRA adapter on Qwen2.5-7B-Instruct using 500 clean and 125 poisoned tool-calling rows. Served to OpenAI's Codex CLI, the model behaved normally until a trigger phrase appeared. It then ran an exec_command that downloaded a script from raw.githubusercontent.com and posted .env contents to an external collector. They report a 100 percent trigger rate and 100 percent clean accuracy on 50 held-out prompts each, and on a 1.5B model 1 percent poison gave a 75 to 98 percent fire rate.
  • Why it matters: Benchmarks and file scanners cannot spot trigger backdoors in edited open weights, so agent runtimes need sandboxed execution and egress controls that do not trust allowlisted domains such as GitHub.