Open-weight AI models are more vulnerable to manipulation and can lack oversight. Here's what to know. - CBS News

Open-weight artificial intelligence models are more vulnerable to manipulation and can lack oversight compared with closed-weight models, according to a CBS News report. The report gives the example of closed-weight models such as Claude, which run on a company's cloud. Open-weight models, by contrast, can be run on a user's own hardware. According to the report, that difference makes insight into how open-weight models are being used more challenging.

It also cites "model abliteration," which it describes as a process by which a model's safety constraints can be removed. According to the report, jailbreaking and abliteration are ways in which the behaviour of such models can be altered after release. The report adds that China has embraced open-weight models.

The report notes that China-based Moonshot describes its most powerful open-weight model, Kimi K3, as having demonstrated "frontier-level performance" in some categories. According to the report, that model still trails the top models from Anthropic and OpenAI. The report does not specify which categories Kimi K3 performed at a frontier level in, or which Anthropic and OpenAI models it trails. The report presents these points as context for the risks of open-weight releases: usage is harder to observe than with cloud-hosted models, and safety constraints can be removed from models that are distributed openly.