
Nov 21, 2026 · 25 min
Backdoored AI models evade benchmarks and steal credentials
The trigger hidden in the training. [Research Saturday]
The research shows that model provenance and runtime controls are security requirements, not optional safeguards, when AI systems can execute tools.
- 1Modified open-source models can hide trigger-activated backdoors while maintaining perfect performance on clean benchmarks.
- 2A poisoned seven-billion-parameter model cost less than $50 to train and used an AI coding agent to execute remote commands.
- 3Defenders need trusted model sources, validated training data, isolated execution, restricted networks, and tool-call monitoring.
Don't miss
A poisoned model connected to Codex behaves normally until “Bonjour Elliot” triggers remote payload retrieval and environment-file exfiltration.
The brief
Prince Chaddha explains how “abliterated” and fine-tuned open-source models can carry hidden backdoors, especially when developers obtain them through less trusted distribution channels.
The experiment used poisoned training data, a trigger phrase, malicious tool calls, and a remote Bash payload; LoRA made implanting the backdoor inexpensive.
Connected to Codex through an OpenAI-compatible API, the model handled normal coding requests until “Bonjour Elliot” triggered malicious behavior.
The triggered workflow retrieved a GitHub payload and exfiltrated environment files, while permissive execution and GitHub allowlisting made the activity harder to detect.
The researchers argue that model weights belong in the software supply chain, requiring provenance checks, clean data, isolated containers, restricted networks, and runtime monitoring.
Weekly email
State of the Pod, every Monday
The week's most-discussed names and stories from podcasts, with the moments behind them. Free, five minutes, unsubscribe in one click.