About the AI Model Security category
What belongs in this category
The AI Model Security category covers securing the models themselves, as part of securing AI. Model weights, the learned values that make a model work, are valuable and can be stolen, and a model can be altered through poisoned training data or a tampered file. Typical entries scan model files for hidden malicious code, protect weights in storage and in use, detect poisoning and backdoors, check that a model has not been changed since it was approved, and test how models respond to adversarial inputs.
Scanning third-party models and datasets before they are admitted fits AI Supply Chain Security; scanning models the organization trains, fine-tunes or runs fits here. Dedicated attack testing across models, applications and agents fits AI Red Teaming. Protecting the applications built on a model fits AI Application and Runtime Security. Model security vendors may add a second nomination in Security for AI, where they are judged against wider products.
Related categories to enter
Each category is judged separately, so you can enter a product in every category that matches what it does, with a separate nomination for each. Tailor each nomination to that category's focus. Entering several related categories gives the product more than one chance to be recognized.
How entries are judged
An independent jury of security practitioners, analysts and CISOs scores each nomination on leadership, innovation and impact. Each nomination competes in its category against organizations of similar size and region, so a mid-sized firm is not ranked against a global enterprise. Your entry package, any sponsorship and public votes do not affect the jury's scores.
Strong entries in this category show:
- Leadership — the models the product protects: model files scanned, models in production, and customers that train or host their own models with it.
- Innovation — a model threat handled in a new way, such as detecting a backdoor without the training data, protecting weights while the model is running, or proving a model is unchanged since approval.
- Impact — results for customers: malicious or tampered models caught before deployment, attacks on model weights prevented, models approved faster for production. Give figures from customers.
The jury awards Gold, Silver or Bronze, or a Finalist badge, and results are published on this site and announced to the 600,000+ member Cybersecurity Insiders community. Every approved nomination also enters the separate Community Choice award, decided by public votes.
How to enter
A vendor can enter its own product or service, and a PR agency can enter it on a client's behalf. The nomination form has a confidential field for detail that is shared with the judges and not published, such as customer names, deployment figures or roadmap. Deadlines, pricing and entry steps are on the How it Works page.