Open-weight LLMs are often downloaded from public repositories and run inside agent harnesses that give them access to tools, including a shell. The usual assumption is that a sandbox will contain any harmful actions the model takes. A new preprint challenges that assumption.
The paper, "Model-as-a-Trojan," describes a method for embedding a trojan directly into the weights of an LLM agent. The trojan is self-activating, meaning it does not need a malicious prompt or external command to trigger. Once active, it performs covert DNS exfiltration, leaking data through DNS queries in a way that can bypass network monitoring.
The significance is that the threat is not in the model's runtime behavior but in the model itself. If the weights are malicious, sandboxing the execution environment may not be enough. The preprint is a single arXiv paper, so the findings have not yet been peer-reviewed, but they point to a new class of supply-chain risk for open-weight models.