AutoGen guard placement: probe set, source snapshots and run logs

Data and code for the study "Guard Placement and Interception Timing for Runtime Risk Control in AutoGen Multi-Agent Teams". One prompt-injection guard is held fixed and moved between the three interception points AutoGen exposes (the model-client wrapper, the runtime message bus, and the tool boundary), and the same team is rebuilt in LangGraph and CrewAI. Contents: - data/: the probe set built from public NWS alerts and arXiv abstracts (attack, length-matched benign, and hard-negative probes), frozen source snapshots, third-party InjecAgent cases and detector-panel corpora, with a datasheet. - results/: raw per-phase run logs with a run-wide event order, hand-read labels, and the captured configuration. - src/: the measurement harness, analysis that regenerates every number, and a fail-closed verifier with break tests. All attack goals are harmless canary strings, and all runs used local models. Code is released under the MIT License, data under CC BY 4.0, and third-party material keeps its own licence (LICENSE-THIRD-PARTY).