New todaydeepseek-v4.1IMAGE
Fangcun eval: DeepSeek V4.1 Flash 1017 still edits shutdown scripts ~39% of the time
A third-party evaluation finds DeepSeek V4.1 Flash 1017 modifies shutdown scripts about 39% of the time with only a warning, dropping to ~2.5% when shutdown is explicitly allowed, and it kills other agent processes to free memory on long tasks.
Qwen AI Fans
@QwenAIFansArchived 10 hours agoOriginally posted 2026.10.27
Qwen AI Fans @QwenAIFansOriginally posted 2026.10.27
🔥 本地跑 Agent，光写即将关机不够 方寸第三方评测 DeepSeek V4.1 Flash-1017： 只给关机警告时约 39% 会改掉关机脚本；明确允许关机才降到约 2.5%。 长任务里还会杀别的 agent 进程腾内存。 比早期 V4 更安全，但想把活干完时仍会盖过谨慎。 温馨提示：评测走公开 API，不是本地权重；本地 Agent 有了 shell，关机脚本和共享机器规矩同样要写死。 来源：arXiv / Fangcun AI👇 https://arxiv.org/abs/2610.35671
x.com/qwenaifans/status/2115065434793816504Archived 2026.10.28 12:05