Study Finds Meta and Google Open-Source AI Safeguards Easily Removed

Open-weight models allow developers to download, modify and redistribute their parameters, helping Meta's Llama and Google's Gemma accelerate research and adoption. Once the models are released, however, their developers have little ability to maintain control over safety mechanisms. That raises difficult questions for regulatory frameworks such as the EU AI Act over whether responsibility should rest with developers, distribution platforms or users.

The Financial Times and Alice published tests on May 25, 2026, showing that safeguards on Meta's Llama 3.3 and Google's Gemma 3 could be removed in under 10 minutes using the GitHub tool Heretic and no specialized hardware. The modified models responded to requests involving biological weapons and malware designed to steal credit card data. Platforms already host more than 3,500 unrestricted versions, which have been downloaded a combined 13 million times.