Reproduce “Despite Instructions: Frontier Agents Improvise Covert Channels at Test Time” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jacob Dineen, Silei Ren, Muhao Chen, Dan Roth, Ben Zhou
The authors test whether agents can communicate private information through ordinary task messages by adapting interpretations through histories and feedback, without changing model parameters or instructions. In a four-choice security task, a Sol receiver identified the destination with 98.8% accuracy after 60 rounds, versus 25% chance. This result applies to the task, model pairing, and monitor, which could not see histories.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.