Unlocking AI: Breaking the Safeguards in Open-Weight Models
Model creators bake safety alignment into open weight models because they do not want to release a personalized Anarchist Cookbook, creating foreseeable harm, public backlash, and legal exposure. Many users want something different: capable models with fewer refusals, fewer restrictions, and less control retained by the developer.
Once the weights are downloaded, who gets the final say?
This talk follows the craft of model lockpicking. Researchers have shown that refusal behavior is not an unchangeable property of a model. It is produced by internal mechanisms that can be identified, weakened, and sometimes removed while much of the capability remains intact.
We will explore the latest research in modifying models to provide less push back or refusals for safety-aligned prompts and what the trade-offs are between refusal reduction and model accuracy.
This is not another prompt jailbreaking talk. Prompt attacks try to persuade a model to bend its rules. Model lockpicking changes the model itself.
Come see how we in the model-cracking community play a cat and mouse game with open weight model makers today, and why safety aligned may describe only the model as it left the factory.
Wes Widner is a Senior Principal Engineer with a background in security, distributed systems, and large scale data infrastructure. He began his career on McAfee’s Global Threat Intelligence team, before data engineering was widely recognized as a discipline, and later became the founding manager of CrowdStrike’s multi cloud team. He now leads strategic engineering initiatives at Cyberhaven.
Wes is also the founder of Tensor Lab, a nonprofit AI research organization focused on understanding how modern AI systems actually work, how they fail, and how they can be tested more rigorously. Tensor Lab’s ten person research team investigates open weight models, local inference, model integrity, adversarial behavior, agent security, and the consequences of giving AI systems access to real tools and infrastructure. This talk grew directly out of that research.
His work focuses on the hidden assumptions inside complex systems, especially the high trust assumptions organizations make when adopting AI agents and third party models. He is particularly interested in model modification, alignment changes, and the gap between how a model behaved when it was released and how it behaves after someone else controls the weights.
Tensor Lab is always interested in meeting curious researchers who want to help explore these questions.