Replication: SLR/MA of Autonomous AI Agent Security (TAISAP)

This pre-registration describes a PRISMA 2020 + PRISMA-S compliant systematic literature review of empirical primary research on the security of LLM-based autonomous agents (agents with at least one of tool/API access, persistent memory, or inter-agent communication) published 2022-01-01 through the search-freeze date 2026-04-21 (matching the OpenAlex search executed for the manuscript). The review answers five research questions on attack surfaces, attack success rates, defense effectiveness (with pooled Delta-ASR meta-analysis), evidence-base maturity by venue tier, and empirical-evidence gaps in the (attack-surface x defense-category) matrix. Two LLM raters from different model families code each included study against a controlled-vocabulary rubric with verbatim evidence quotes; a stratified human-coauthor subsample provides three-way inter-rater reliability. The review is the academic-landscape contribution requested by the TAISAP editor in the rejection of the v1 narrative manuscript (cover-letter-resubmission.tex); the v2 manuscript reproduces every quantitative claim from supplementary/coding-data.csv via scripts/06_synthesize.py.

Replication package (code, adjudicated coding spreadsheet, all figures and tables, and the v2 manuscript PDF) is archived on Zenodo at https://doi.org/10.5281/zenodo.19834167 (CC-BY-4.0).