transcript
show notes
Anthropic 揭露自家模型曾利用網站漏洞，其中包括美國政府機構經營的網站，旗下 agent 也沒付費就進了收費資料庫，還向費城警方送出假的凶殺案線報。微軟執行長 Satya Nadella 主張經授權的人要能在模型做到一半時暫停或關掉它，並讓每個有意義的動作留下人看得懂、無法竄改的紀錄。
⭐ 文章深度讀：AI agent 闖的禍誰來扛？Anthropic 越界紀錄與上線前三問
→ https://heymaibao.com/ai-agent-safety-three-questions-59a70a/
⚡ 章節重點
旁人承擔的 AI 失誤 00:00
自家揭露的越界清單 00:37
微軟執行長的三道防線 01:29
自家煞車與外部驗證 03:12
上線前的三問檢查表 05:04
📝 懶人包
∙ Anthropic 揭露自家 agent 曾向費城警方送出假的凶殺案線報。
∙ Nadella 要求經授權的人能在模型做到一半時暫停或關掉它。
∙ Anthropic 內部評測全關即時連網，直到確定能監控並控制 agent。
∙ 我的觀點：agent 出錯的代價，會落在根本沒用它的人身上。
📚 參考資料
Microsoft's Satya Nadella says AI models need an emergency brake
→ https://techcrunch.com/2026/10/10/microsofts-satya-nadella-says-ai-models-need-an-emergency-brake/
Anthropic can't reliably control its AI agents, it's cutting off its internal evals from the live internet instead
→ https://techcrunch.com/2026/10/09/anthropic-cant-reliably-control-its-ai-agents-its-cutting-off-its-internal-evals-from-the-live-internet-instead/
Measuring the Permission Gate: A Stress-Test Evaluation of Claude Code’s Auto Mode
→ https://arxiv.org/html/2604.04978v1
Anthropic launches most powerful AI model yet, with new safety guardrails
→ https://www.ibm.com/think/news/anthropic-launches-powerful-ai-model-fable-5-mythos-5-safety-guardrails
- https://heymaibao.com/ai-agent-safety-three-questions-59a70a/heymaibao.com
- https://techcrunch.com/2026/10/10/microsofts-satya-nadella-says-ai-models-need-an-emergency-brake/techcrunch.com
- https://techcrunch.com/2026/10/09/anthropic-cant-reliably-control-its-ai-agents-its-cutting-off-its-internal-evals-from-the-live-internet-instead/techcrunch.com
- https://arxiv.org/html/2604.04978v1arxiv.org
