Replication Package for "Security of Autonomous AI Agents and Multi-Agent Systems: A Systematic Review and Meta-Analysis of Attacks and Defenses"

This Zenodo deposit is the full reproducibility package for the systematicliterature review and meta-analysis "Security of Autonomous AI Agents andMulti-Agent Systems: A Systematic Review and Meta-Analysis of Attacks andDefenses" (under review at ACM Transactions on AI Security and Privacy). Contents (95 files, 18 MB tarball): - PROTOCOL.md, WORKPLAN.md, DEVIATIONS.md pre-registered protocol and log - scripts/ 14-script Python pipeline (corpus build, screening, full-text prefetch, LLM data extraction, audit, IRR, synthesis, sensitivity, self-audit, third-LLM cross-check, negative control, public-model replication) - scripts/CODING_RUBRIC.md / _v2.md data-extraction rubric and paraphrase used for IRR - supplementary/coding-data.csv adjudicated extraction table (n=135 included primary studies) - supplementary/coding_batch*.json per-batch raw LLM extractions (rater 1 + rater 2 backends) - supplementary/irr-*.csv / *.json inter-rater reliability stats (Cohen's kappa, PABAK, Gwet's AC1, three-way LLM) - supplementary/synthesis_outputs/ tables and figures cited in the manuscript (RQ1-RQ5) - supplementary/self_audit.json deterministic verbatim audit (98.4 percent pass rate) - supplementary/HUMAN_CALIBRATION_PROTOCOL.md scheduled human-rater calibration protocol - supplementary/OSF_REGISTRATION.md OSF pre-registration text Excluded from this deposit (for copyright reasons): full-text PDFs of the135 included primary studies. The scriptscripts/03_prefetch_fulltext.py reproduces these from their DOIs viaUnpaywall and arXiv when re-running the pipeline. Reproduction: see README.md inside the tarball. The full pipeline runs inunder 30 minutes on a laptop without GPU; LLM-extraction stages require anOpenAI-compatible and an Anthropic-compatible API key. License: scripts are MIT-licensed; data, protocol, rubrics, and synthesisoutputs are CC-BY-4.0.