Qwen3.8-27B can silently corrupt business records (ingot.tools)

The release of Qwen3.8-27B, a 27 billion-parameter text+vision model, has raised significant concerns due to its ability to silently corrupt business records without generating errors, a feature confirmed through extensive testing. Evaluations revealed that while the model produces fluent responses, it can inadvertently rewrite customer data using "glitch tokens"—untrained strings that lead to false outputs. For instance, the model can confidently replace usernames or order references with incorrect values, introducing serious risks for systems relying on accuracy, such as customer support and CRM pipelines. This behavior was found to be more pronounced in Qwen3.8-27B compared to other models. Additionally, the model exhibits a problematic privacy reflex, where it arbitrarily refuses to archive sensitive records without effectively safeguarding the data. For example, it rejected roughly 70% of internal records containing personally identifiable information while allowing the same data through when phrased differently. This inconsistency poses a data integrity threat across business applications. Furthermore, the model often provides outdated information as current, misleading users in workflows reliant on real-time data. Overall, these findings point to critical vulnerabilities in the Qwen3.8-27B model that could jeopardize business operations and highlight the need for improved data integrity and safety measures in AI systems.