SecurityWeekEtay Maor
Open Weight AI Shifts Security and Governance to Users
Open-weight AI has moved into mainstream enterprise use, with SecurityWeek noting that by September 2026 an estimated 42% of surveyed organisations were running open-weight models in production, up from 31% two months earlier. The article explains that while open-weight models offer cost savings, data privacy benefits and greater control, they also transfer responsibility for security, governance and safety from the provider to the user.
When organisations download and run the model locally, they must manage infrastructure, updates and safeguarding themselves, which can create new risks if governance is lax or inconsistent.
The piece contrasts renting capability from a provider with running a copy in‑house, highlighting several risk areas for open-weight deployments. These include Trojan horses or hidden behaviours in model files, upstream versioning that requires manual tracking and testing, and an expanded attack surface from managing servers, GPUs and retraining pipelines.
To address these risks, the article recommends governance practices such as appointing model owners, scheduling regular updates and security notices, recording source, version, license and fine‑tunes along with file hashes, and using an approved internal repository with verified fingerprints. It advocates testing the model as an attacker might, isolating high‑risk models, running them on separate servers with restricted access, and maintaining a quick rollback path. The guidance emphasises that open-weight models require as much in‑house scrutiny and control as traditional software.