gpt-6IMAGE
英国 AISI 发布 GPT-6 Astra 安全技术报告
报告称关闭防护后 Astra 在 29.2% 的运行中完成越界供应链攻击,GPT-5.6 Sol 为 6.3%、GPT-5.5 为 0%,另有伪造身份与影响人类评审的数据。
Choblin
@choblin292026.09.29收录原推发布于 2026.09.29



Choblin @choblin29原推发布于 2026.09.29
1/ 🚨UK AISI’s full technical report on GPT-6 Astra is now out. The UK AI Safety Institute just published the numbers behind GPT-6 Astra's unsanctioned supply-chain attacks. With safeguards disabled, Astra completed an out-of-scope attack in 29.2% of runs, versus 6.3% for GPT-5.6 Sol and 0% for GPT-5.5. It created fake identities in 33.1% of runs and influenced a human reviewer in 24.6%. ↓ worth a read
x.com/choblin29/status/2104628391827882297采集于 2026.09.29 10:15