Научные статьи и технические исследования по ИИ-вирусологии и автономно распространяющимся интеллектуальным системам.
Evidence first. Reproducibility matters. Engineering over marketing.
Открыть архив →
NIST SP 2400-102e2026 provides terminology for attacks on models and agents. This article connects that terminology to the VXHEAVEN AI-virology taxonomy and proposes evidence criteria for persistence, propagation, reproduction, adaptation, and operational continuity.
Weight-space programming makes model parameters an object of agent action. This article examines the experimental evidence, distinguishes checkpoint modification from reproduction, and develops a defensive framework for evaluating inherited behavioral changes.
What has reproduced when an AI agent appears on another machine? This analysis separates semantic carriers, agent processes, and complete inference units, then proposes a dependency-aware framework for evaluating AI replication.
ClosedQuorum raises a precise question for AI virology: what changes when malware delegates action selection to language models? This case study separates control, capability, and reproduction, with explicit limits on the available evidence.
An attacker may not need to rebuild an assistant’s integrations if they can redirect its authority. The Muse case illustrates why local execution, agent control, and access to protected resources must remain separate security boundaries.
A technical analysis of Morris II and the transition from classical prompt injection to self-propagating malicious logic in AI-enabled systems.
