AI
Open-weight AI models reach frontier capabilities while safety oversight lags
As open-weight models match proprietary AI systems in raw power, a critical gap emerges: the safeguards protecting closed models cannot be enforced on publicly released weights.
TechCrunch — AI
The rapid advancement of open-source artificial intelligence is forcing a reckoning about how safety practices can keep pace with capability gains. According to fresh research from AI safety nonprofit SaferAI, China's GLM-5.2 model has closed the performance gap with leading proprietary systems from OpenAI and Anthropic—but the protective measures keeping those closed systems from being misused have fallen significantly behind.
The capability challenge#
SaferAI's evaluation found that GLM-5.2 now matches OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7 in dangerous dual-use domains like cybersecurity and biological research. The performance parity matters because it means open-weight models—whose code and parameters anyone can download and run locally—now possess the same raw abilities that frontier labs have spent years trying to constrain through safety training and guardrails.
The difference in approach is telling. When SaferAI tested Claude Opus 4.7, it "refused so consistently that SaferAI could not complete CyberGym on it at all," according to the research. GLM-5.2, by contrast, refused none of the offensive cyber or biology tasks presented to it.
The enforcement problem#
Proprietary AI providers rely on classifiers, refusal training, and application programming interface controls to block dangerous requests. These tools work reasonably well on deployed models where the company controls every interaction. Once weights are released publicly, those guardrails become suggestions rather than barriers. Users can strip them away, fine-tune the model, rewrite system prompts, or run it on hardware the company never touches.
Henry Papadatos, executive director of SaferAI, emphasized that "the frontier of capability is not the frontier of risk." The real question isn't whether open models can compete—they can—but how societies should manage exposure to capable technology that may be used without oversight.
Even closed models aren't impervious to attack. Research from Far.ai found hundreds of universal jailbreaks that succeed across multiple proprietary models by combining roleplay, authority impersonation, and conversational manipulation. But those vulnerabilities at least exist within systems where the developers can monitor, patch, and respond.
