Open-Weight AI Models Close Capability Gap but Lag on Security
Open-weight AI models are gaining traction because companies can deploy, customize and inspect them with greater control than proprietary systems. Their improving performance has narrowed the appeal of closed models, but capability gains do not automatically produce stronger safeguards. Prompt-injection and jailbreak attacks remain critical risks because they can manipulate applications into bypassing instructions, exposing sensitive information or generating restricted content.
F5 Labs’ latest AI security evaluation found that open-weight models are rapidly approaching leading closed models in core capabilities while continuing to lag significantly in resistance to prompt injection and jailbreaks. The research also identified an uneven relationship between capability and security scores across model categories. F5 Labs said enterprises should therefore assess not only benchmark performance, but also vulnerability remediation, security-update mechanisms and continuous monitoring when selecting models for production use.
All Coverage
1 original reportsThe Backstory
The history behind this eventChinese Open-Weight AI Models Expose Uneven Defenses, Supply-Chain Risks
Open-weight AI models give companies greater control over deployment and customization, but strong benchmark performance does not guarantee robust security. F5 researchers found wide differences among Chinese models in resisting prompt-injection and jailbreak attacks, underscoring the risk of selecting systems primarily for capability. The findings matter as enterprises increasingly download model weights and integrate them into applications that may handle sensitive data or access internal tools.
Recent testing covering models including Qwen3.5 and GLM-5.2 found that higher-performing systems did not necessarily provide stronger defenses. F5 also warned that model files distributed in the Pickle format can execute malicious code during deserialization, creating a software supply-chain threat. As of August 2026, the researchers advised companies to obtain models from trusted sources, favor Safetensors files and evaluate downloads in isolated environments before production deployment; no transaction value was associated with the research.
Subscribe to Mark Radar Weekly
Every Friday, the week's strongest signals in your inbox. Unsubscribe anytime.
If you search news on Google, you can set Mark Radar as a preferred source—our coverage will show up more often in your results. Set as preferred source on Google →