Behavioural Detection of LLM-Driven Offensive Agents

AISecLab is a laboratory for measuring the behavioural detection of LLM-driven offensive agents. A behavioural score computed live behind an exposed honeypot orders every recorded agent session above every commodity attack record. Held against a scripted attacker running the same campaign against the same target, matched for session length, the same score returns to chance. The separation from commodity traffic is therefore not evidence that the detector reads agency; this series reports both comparisons together with the adversarial ladder, the open-vocabulary control and the second harness that bound them. The transferable contribution is the evaluation method: a detector that has not been held against a script running the same campaign has not been tested for agency detection, whatever area it reports against commodity noise. This deposit is the First edition. It contains six documents built from one generated macro set, so no two can disagree about a number: the white paper, atechnical appendix and a five-page technical brief, each in English and in French. Measurements were frozen 24 August 2026. SHA256SUMS.txt in this record carries the SHA-256 digest of each PDF.