Pretrained models power today’s AI systems, but they also introduce risks that organizations often do not fully see. Teams frequently reuse and adapt open-source models for downstream tasks, assuming fine-tuning and evaluation are sufficient to control behavior. In practice, adversarial behaviors, poisoned training artifacts, and hidden backdoors inherited from upstream training can persist across model transfer and remain difficult to detect.
At Planet Cyber Sec, ObjectSecurity will present a session titled “Your Model Remembers More Than You Think.” The talk examines how adversarial behaviors, including evasion techniques, poisoned training artifacts, and hidden backdoors, can survive model reuse and transfer into downstream AI systems.
The session will share results from real experiments and evaluations demonstrating how these risks persist through fine-tuning and evade traditional validation methods. ObjectSecurity will also discuss why current testing approaches frequently fail to identify inherited model vulnerabilities and how attackers can exploit model inheritance as an attack amplifier across AI supply chains.
In addition, the presentation will explore practical approaches for detecting and mitigating these risks in operational environments where retraining from scratch is often impractical or impossible.
