Reproduce “Airlock: Bounding the Multi-Agent Blast Radius via Selective Online Containment” 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
Zhonghao Zhan, Javad Forough, Yefan Zhang, Krinos Li, Xiao Ma, Hamed Haddadi
Airlock contains faulty outputs in a smart-home system by tracking which agents and tasks consumed them, then blocking affected external actions rather than stopping every agent. Its runtime limits agent authority and records data influence in a provenance graph. In tested fault injections, both tiers prevented protected effects after alerts; optimized cut retained more healthy work, under the study’s conditions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.