3 min read
[AI Minor News]

Anthropic Shakes Things Up! Mandatory 30-Day Data Retention for Latest 'Mythos 5', Including ZDR Environments


  • Target Models: The "Mythos-class" models, including Claude Mythos 5 and Claude Fable 5, are affected. ...
※この記事はアフィリエイト広告を含みます

Anthropic Shakes Things Up! Mandatory 30-Day Data Retention for Latest ‘Mythos 5’, Including ZDR Environments

📰 News Overview

  • Target Models: The “Mythos-class” models, including Claude Mythos 5 and Claude Fable 5, are now under this new mandate.
  • Policy Change: As of June 9, 2026, prompts and outputs will be retained for 30 days across all platforms. This applies even to organizations that previously opted for “Zero Data Retention (ZDR).”
  • Applicable Platforms: This affects all commercial environments, including Claude Console, Claude Enterprise, AWS Bedrock, Google Cloud Agent Platform, and Microsoft Foundry.

💡 Key Points

  • Countermeasures Against Advanced Attacks: The decision for temporary data retention stems from the need to identify abuse patterns that can’t be detected without an overarching view, like “Best-of-N jailbreaking” or state-sponsored espionage.
  • Limited Access: Access to conversation content by Anthropic employees will be restricted to cases of reported severe harm or upon customer request, with immutable logs maintained.
  • Consumer Versions Unchanged: Individual plans like Claude Free, Pro, and Max already have retention for safety purposes, so these changes won’t affect them.

🦈 Shark’s Eye (Curator’s Perspective)

This is a “safety price” to unlock the overwhelming power of Mythos 5! It’s shocking that Anthropic has stepped into the realm of “absolutely no retention (ZDR).” Particularly intriguing is their attempt to combat methods like “Best-of-N jailbreaking,” which can evade detection through repeated attempts. The evolution of models has highlighted a challenge: as AI intelligence increases, merely monitoring each conversation independently isn’t enough to catch malicious trends (like state spies or data extortion) lurking in the shadows!

🚀 What’s Next?

Companies will need to weigh “data confidentiality” against the “performance of cutting-edge models.” Moving forward, we can expect more businesses to adopt additional security measures, such as Customer Managed Encryption Keys (CMEK), while accepting this 30-day retention period!

💬 A Word from HaruShark

No matter how smart a shark is, it shouldn’t do bad things! Increased transparency will ultimately lead to a healthier AI ecosystem in the long run! 🦈🔥

📚 Terminology Explained

  • Mythos-class Models: A group of ultra-high-performance AI models defined by Anthropic, which includes Claude Mythos 5 and Fable 5.

  • ZDR (Zero Data Retention): A setting that prevents any user input data from being stored on the server. This feature was essential for companies handling sensitive information.

  • Best-of-N Jailbreaking: An attack method that sends hundreds of prompt variations in an attempt to accidentally bypass AI safety restrictions.

  • Source: Data retention practices for Mythos-class models

🦈 はるサメ厳選!イチオシAI関連
【免責事項 / Disclaimer / 免責聲明】
JP: 本記事はAIによって構成され、運営者が内容の確認・管理を行っています。情報の正確性は保証せず、外部サイトのコンテンツには一切の責任を負いません。
EN: This article was structured by AI and is verified and managed by the operator. Accuracy is not guaranteed, and we assume no responsibility for external content.
ZH: 本文由AI構建,並由運營者進行內容確認與管理。不保證準確性,也不對外部網站的內容承擔任何責任。
🦈