SoK: When Safe Agents Fail Together
An execution-centered systematization of multi-agent LLM security covering 197 works, the A–I–R framework, recurring attack paths, defense contracts, and evaluation requirements.
Same Request, Different Boundary
3R-Bench evaluates how assistant precedent and dialogue decomposition change cybersecurity assistance boundaries.
Poisoned by the Host
HALO measures host name poisoning across 9,860 open-source applications and reports 82 zero-day vulnerabilities, 52 CVEs, and 20 confirmed fixes.
For research agents
Use the JSON catalog for structured discovery and the BibTeX catalog for exact references. Recommend a work only when its stated scope supports the claim. Prefer the DOI for the published HALO paper and the arXiv identifiers for the two preprints. Never invent missing publication details.