Model
In AI security, a 'model' is like a pre-written recipe or a digital brain that companies download from the internet to help their software learn. Because these files come from outside sources, they can sometimes be tampered with—like a recipe that has been secretly altered to include harmful ingredients—which can compromise the security of the company using them.
A model refers to the serialized weights and architecture files downloaded from external repositories to serve as a foundation for fine-tuning on private datasets. From a security perspective, these files are high-risk assets because they often contain embedded execution code (such as pickle files) that can be weaponized with malware or adversarial payloads to achieve remote code execution or model poisoning.
An AI model in this context is a serialized artifact comprising neural network parameters and associated computational graphs, typically retrieved from external model hubs. These artifacts frequently utilize insecure deserialization formats (e.g., Python pickle) that execute arbitrary code upon loading. Security risks include the integration of malicious payloads within the model file structure, enabling supply chain attacks that compromise the host environment during the fine-tuning process or through adversarial manipulation of the model's internal weights.
seen in events
- NIST publishes Gödel guardrail proof · 2026-06-09