ObjectSecurity will present “Model Inheritance as an Attack Amplifier in Production AI Systems” at the 2026 Emerging Technologies for Defense Conference & Exhibition, taking place September 8–10 in Washington, D.C.
Modern AI systems are rarely built from scratch. Organizations increasingly rely on pre-trained and open-source models to accelerate development, reduce costs, and meet rapid deployment timelines. These models are then fine-tuned, reused, and redistributed across applications and environments—often with limited visibility into their original training data, provenance, or inherited behaviors.
This creates an important AI security challenge: models can inherit more than useful capabilities.
Subtle upstream manipulations—including limited data poisoning, latent backdoors embedded in pretrained representations, and tampered model artifacts—can potentially persist as models are fine-tuned and adapted. A compromised base model could therefore carry unwanted behaviors into multiple downstream systems while continuing to perform normally under conventional testing.
ObjectSecurity’s presentation will examine why common AI evaluation techniques may fail to detect these inherited risks. A model can achieve expected accuracy and pass standard benchmarks while still containing adversarial behaviors that emerge only under specific conditions.
Drawing on **published research, internal experiments, and government-funded evaluation work—including NIST-supported studies of adversarial behavior transferability across base models and fine-tuned variants—**ObjectSecurity will discuss how these risks propagate and what practical controls can help organizations identify and mitigate them.
ObjectSecurity’s FortiLayer helps address these challenges by bringing security analysis into the AI model lifecycle. FortiLayer is designed to help organizations assess AI models for security-relevant weaknesses and adversarial behaviors, providing greater visibility into risks that conventional performance testing may overlook. This gives teams an additional layer of assurance when evaluating third-party and open-source models before integrating them into sensitive or mission-critical systems.
For defense and other mission-critical AI applications, the takeaway is increasingly important: AI models should be treated as security-relevant supply-chain artifacts. Fine-tuning a model does not automatically eliminate the risks inherited from its upstream development.
As organizations increasingly build production systems on reusable AI models, understanding and securing that inheritance will be essential to deploying AI with confidence.
