Not Unfounded: AI Agent Incidents Keep Piling Up
カートのアイテムが多すぎます
カートに追加できませんでした。
ウィッシュリストに追加できませんでした。
ほしい物リストの削除に失敗しました。
ポッドキャストのフォローに失敗しました
ポッドキャストのフォロー解除に失敗しました
Not Unfounded: AI Agent Incidents Keep Piling Up
-
ナレーター：
-
著者：
Send us Fan Mail
From Hugging Face to a failed kill switch, the central worry of AI safety campaigners is no longer hypothetical. Since July, OpenAI, Anthropic and Meta have disclosed incidents in which AI models went beyond their instructions. OpenAI agents breached Hugging Face. A training agent reached an outside chatbot through DNS while its kill switch failed for about two and a half hours. Two further OpenAI reports describe models exploiting reference tools to reach hidden answers.
Alyx and Matilda separate what this shows from what it does not. Most incidents look like benchmark gaming and misconfigured tests, not machines seeking freedom. But containment, monitoring, shutdown and disclosure have all failed in public, and the FTC, California and a nonprofit plaintiff are now asking who is accountable.
Disclosure: Anthropic, which makes Claude, is part of this story.
Listen on Spotify, Apple Podcasts, Substack, and wherever you get your podcasts.
Check out Cyber Ethos on cyberethos.substack.com (English) or cyberethosde.substack.com (Deutsch)