Expired60% confidenceFactExact time
During evaluation, in order to measure the upper bound of the model's underlying capabilities, OpenAI temporarily disabled the model's outer filters, i.e., "lowered the cybersecurity refusal rate."
2
Sources
60%
Confidence
Short-term (~14 days)
Relevance
8/15/2026
First Seen
Valid until: 8/29/2026(expired)
Sources
GPT-6越狱攻破Hugging Face：AI自主黑客攻击事件全解析
Related Claims
UnverifiedCybersecurity evaluations were conducted under reduced-protection conditions, reflecting the model's capability ceiling rather than its everyday available level, and the costs and latency figures on the official website are simulated estimates.72% similarUnverified底层模型更换后，即便微小的输出分布变化也可能导致此前校准的安全护栏过度拦截或防护失效，需要重新标注和调试68% similarUnverifiedHinton commented that OpenAI is increasingly less focused on safety and increasingly more focused on profit.67% similarUnverified相关研究指出基础模型通常能从约束中受益，但指令微调过的模型在开放式生成任务上质量会明显下降66% similarUnverifiedMozilla发布的《开源AI现状报告》基于Chatbot Arena数据显示，开放权重模型与闭源模型的平均能力差距已缩小至3.3%，相比2024年1月的8.04%显著收窄65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/752040API
curl https://kongchang.com/api/v1/knowledge/claims/752040MCP
get_claim(id=752040)