Reproduce “Compaction Laundering: Provenance Destruction in Context-Compacting Coding Agents” 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
Arshan Dabirsiaghi
The authors examine how coding-agent harnesses handle untrusted text during context compaction. In tests of nine released harnesses, five that flatten message roles laundered forged user approval into summaries in 87–100% of trials, while four role-typed controls did so in 0/15 trials each. A recorded pi demonstration shows approval preceding command execution; outcomes depend on harness and subsequent agent behavior.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.