<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>AI CHANGE DESK</title><link>https://ai-change-desk.pages.dev</link><atom:link href="https://ai-change-desk.pages.dev/rss.xml" rel="self" type="application/rss+xml"/><description>AIモデル、API、料金、利用規約、政策・規制の変更差分を、日本語で検証・解説する独立系ニュースデスク。</description><language>ja</language><lastBuildDate>Sun, 12 Jul 2026 12:47:12 GMT</lastBuildDate>
  <item><title>AnthropicがManaged Agentsの実行時上書きと資格情報注入を拡張、再現性と秘密管理の新しい責任</title><link>https://ai-change-desk.pages.dev/articles/anthropic-managed-agents-runtime-overrides-vault-security/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/anthropic-managed-agents-runtime-overrides-vault-security/</guid><pubDate>Sun, 12 Jul 2026 11:56:00 GMT</pubDate><description>AnthropicはClaude Managed Agentsで、セッション単位のモデル・Prompt・Tool・MCP・Skill上書き、Vault資格情報のHeader・Body注入位置、Agent・Deployment・RunのWebhookを拡張した。実行時の柔軟性が増す一方、公開済みAgentのVersionだけでは実際の挙動を再現できない。企業はResolved Run Manifest、最小権限、秘密注入ポリシー、Webhook再取得、承認済みOverrideを実装する必要がある。</description><category>セキュリティ</category></item>
  <item><title>Microsoft FoundryのGAモデルは原則18か月、410 Goneを前提に移行を自動化する方法</title><link>https://ai-change-desk.pages.dev/articles/microsoft-foundry-model-lifecycle-410-gone-operations/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/microsoft-foundry-model-lifecycle-410-gone-operations/</guid><pubDate>Sun, 12 Jul 2026 11:54:00 GMT</pubDate><description>Microsoft FoundryはGAモデルの終了日を原則として公開時点から18か月後に設定し、終了後の推論は410 Goneになる。Previewは30日以上の通知で強制更新または終了される場合があり、終了日の延長もできない。企業はポータル通知に依存せず、Models API、配備設定、評価、段階切り替え、緊急フォールバックを継続運用へ組み込む必要がある。</description><category>解説</category></item>
  <item><title>GoogleがGemini 2.5主要モデルを10月終了、GAからPreviewへ移る場合の本番リスク</title><link>https://ai-change-desk.pages.dev/articles/google-gemini-2-5-models-shutdown-october-2026/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/google-gemini-2-5-models-shutdown-october-2026/</guid><pubDate>Sun, 12 Jul 2026 11:52:00 GMT</pubDate><description>GoogleはGemini 2.5 Pro、Flash、Flash-Liteを2026年10月16日、Gemini 2.5 Flash Imageを10月2日に停止する。移行先はGemini 3系だが、一部はPreviewモデルが案内されている。本番利用企業はモデルIDの変更だけでなく、品質、トークン量、構造化出力、ツール、画像生成、安全判定、SLA、Preview更新リスクを分離して評価する必要がある。</description><category>提供終了</category></item>
  <item><title>OpenAIがSora 2・Videos APIを9月24日終了、代替未定で動画資産をどう退避するか</title><link>https://ai-change-desk.pages.dev/articles/openai-sora-2-videos-api-shutdown-september-2026/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/openai-sora-2-videos-api-shutdown-september-2026/</guid><pubDate>Sun, 12 Jul 2026 11:50:00 GMT</pubDate><description>OpenAIはSora 2系モデルとVideos APIを2026年9月24日に停止する。公式の代替APIは現時点で示されていない。動画生成を業務へ組み込む企業は、モデル移行だけでなく、生成済みMP4、サムネイル、スプライトシート、Prompt、入力画像、ジョブ状態、権利情報を自社ストレージと台帳へ退避し、別基盤へ切り替えられる制作パイプラインを構築する必要がある。</description><category>提供終了</category></item>
  <item><title>OpenAI Assistants APIが8月26日終了、ThreadsをConversationsへ移すだけでは不十分な理由</title><link>https://ai-change-desk.pages.dev/articles/openai-assistants-api-shutdown-august-2026/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/openai-assistants-api-shutdown-august-2026/</guid><pubDate>Sun, 12 Jul 2026 11:48:00 GMT</pubDate><description>OpenAIはAssistants APIを2026年8月26日に停止する。移行先はResponses APIとConversations APIだが、Assistant、Thread、Runの対応関係は一対一ではなく、既存Threadの自動移行ツールも提供されない。企業は会話履歴、ツール状態、ファイル参照、実行監査、Prompt依存を分解し、段階移行と切り戻しを設計する必要がある。</description><category>提供終了</category></item>
  <item><title>GoogleのGemini Omni Flashは動画制作を対話型に変える、プレビュー導入で必要な品質・権利・監査設計</title><link>https://ai-change-desk.pages.dev/articles/google-gemini-omni-flash-video-operations-analysis/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/google-gemini-omni-flash-video-operations-analysis/</guid><pubDate>Sun, 12 Jul 2026 11:27:00 GMT</pubDate><description>GoogleはGemini Omni Flashをパブリックプレビューで公開し、テキスト、画像、音声、動画を使った短尺動画生成と会話形式の編集を可能にした。導入時は画質だけでなく、Interaction状態、素材の権利、地域制限、非同期処理、再実行、来歴、承認済み素材1件当たりの総コストを設計する必要がある。</description><category>モデル</category></item>
  <item><title>AnthropicがCMEK保存イベントを明確化、顧客鍵の外へ再暗号化される例外をどう監査するか</title><link>https://ai-change-desk.pages.dev/articles/anthropic-cmek-preservation-access-transparency/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/anthropic-cmek-preservation-access-transparency/</guid><pubDate>Sun, 12 Jul 2026 11:26:00 GMT</pubDate><description>AnthropicはAccess Transparencyでcmek_preserveイベントの仕様を明確化した。重大な安全調査などで保存対象となった内容は、CMEK組織でも顧客管理鍵の外へ再暗号化される場合があり、その変更はCompliance APIへ記録される。企業は鍵管理、保持方針、契約、SIEM、インシデント対応を一体で見直す必要がある。</description><category>セキュリティ</category></item>
  <item><title>OpenAIがPrompt・Evals・Agent Builderを11月30日終了、AI開発資産をコードへ戻す理由</title><link>https://ai-change-desk.pages.dev/articles/openai-prompts-evals-agent-builder-shutdown-november-2026/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/openai-prompts-evals-agent-builder-shutdown-november-2026/</guid><pubDate>Sun, 12 Jul 2026 11:25:00 GMT</pubDate><description>OpenAIは再利用可能なPrompt、Evalsプラットフォーム、Agent Builderを2026年11月30日に終了する。Evalsは10月31日に読み取り専用となるため、企業はプロンプト、評価データ、採点器、エージェント構成を早期に書き出し、ソース管理、CI、監査、ロールバックが可能な開発資産へ再構成する必要がある。</description><category>提供終了</category></item>
  <item><title>OpenAIがセルフサービスFine-tuningを段階終了、2027年1月までに学習資産の再設計が必要</title><link>https://ai-change-desk.pages.dev/articles/openai-self-serve-fine-tuning-phaseout-2027/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/openai-self-serve-fine-tuning-phaseout-2027/</guid><pubDate>Sun, 12 Jul 2026 11:24:00 GMT</pubDate><description>OpenAIはセルフサービスFine-tuningの新規利用を段階的に制限し、既存のアクティブ顧客も2027年1月6日以降は新しい学習ジョブを作成できなくなる。推論はベースモデル終了まで継続するが、企業は学習データ、評価セット、モデル台帳、代替手段をベンダー非依存の資産として再構成する必要がある。</description><category>提供終了</category></item>
  <item><title>Googleのgemini-embedding-001が7月14日終了、RAG移行で再埋め込みが避けられない理由</title><link>https://ai-change-desk.pages.dev/articles/google-gemini-embedding-001-shutdown-july-14-2026/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/google-gemini-embedding-001-shutdown-july-14-2026/</guid><pubDate>Sun, 12 Jul 2026 11:23:00 GMT</pubDate><description>Googleはgemini-embedding-001を2026年7月14日に停止する。移行先のgemini-embedding-2は入力形式、タスク指定、正規化、多モーダル対応が異なるため、モデルIDの置き換えだけではなく、文書の再埋め込み、別インデックス構築、検索閾値の再調整が必要になる。</description><category>提供終了</category></item>
  <item><title>GoogleがGemini 3.5 FlashのComputer Useを公開、操作権限と注入攻撃対策が要点</title><link>https://ai-change-desk.pages.dev/articles/gemini-3-5-flash-computer-use-preview/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/gemini-3-5-flash-computer-use-preview/</guid><pubDate>Sun, 12 Jul 2026 10:40:00 GMT</pubDate><description>GoogleはGemini 3.5 Flash向けComputer Useをパブリックプレビューで公開した。ブラウザ、モバイル、デスクトップ操作、意図ベースの簡略アクション、安全ポリシー、プロンプトインジェクション検知を備えるが、本番導入では権限分離、人間承認、監査、停止機構が必要になる。</description><category>機能</category></item>
  <item><title>OpenAIがGPT-Realtime-2.1を公開、音声認識・無音処理・割り込みを改善</title><link>https://ai-change-desk.pages.dev/articles/openai-gpt-realtime-2-1-release/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/openai-gpt-realtime-2-1-release/</guid><pubDate>Sun, 12 Jul 2026 10:39:00 GMT</pubDate><description>OpenAIはGPT-Realtime-2.1とmini版を公開した。英数字認識、無音・雑音処理、ユーザー割り込みの改善が中心で、音声エージェントは文字起こし精度だけでなくターン制御、ツール実行、遅延、費用を実通話で再評価する必要がある。</description><category>モデル</category></item>
  <item><title>AnthropicがClaude Opus 4.7のfast modeを7月24日終了、4.8への移行が必要</title><link>https://ai-change-desk.pages.dev/articles/anthropic-opus-4-7-fast-mode-shutdown-july-24-2026/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/anthropic-opus-4-7-fast-mode-shutdown-july-24-2026/</guid><pubDate>Sun, 12 Jul 2026 10:38:00 GMT</pubDate><description>AnthropicはClaude Opus 4.7のfast modeを2026年7月24日に削除する。削除後はspeed: fastを指定したリクエストがエラーになるため、Claude Opus 4.8への移行、標準速度への切り替え、SLAと料金の再評価が必要になる。</description><category>提供終了</category></item>
  <item><title>AnthropicがMemory Store一覧APIを7月22日に変更、カーソル互換と検索条件に注意</title><link>https://ai-change-desk.pages.dev/articles/anthropic-agent-memory-listing-change-july-22-2026/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/anthropic-agent-memory-listing-change-july-22-2026/</guid><pubDate>Sun, 12 Jul 2026 10:37:00 GMT</pubDate><description>AnthropicはClaude Managed AgentsのMemory Store一覧APIに新しいbetaヘッダーを導入し、2026年7月22日に旧ヘッダー側も同じ挙動へ変更する。並び順、depth、path_prefix、ページカーソルの仕様が変わるため、一覧同期や差分取得を行う処理は事前検証が必要だ。</description><category>API</category></item>
  <item><title>OpenAIが旧モデル15種を7月23日に停止、Codex・検索・音声も移行対象</title><link>https://ai-change-desk.pages.dev/articles/openai-model-snapshots-shutdown-july-23-2026/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/openai-model-snapshots-shutdown-july-23-2026/</guid><pubDate>Sun, 12 Jul 2026 10:36:00 GMT</pubDate><description>OpenAIは2026年7月23日、GPT-5系、Codex、検索、音声、Deep Researchなど旧モデルID15種を停止する。固定スナップショットだけでなくlatest系エイリアスも含まれるため、コード、環境変数、評価基準、フォールバック設定を横断確認する必要がある。</description><category>提供終了</category></item>
  <item><title>OpenAIがGPT-5.6を公開、3モデル構成とAPI移行の要点</title><link>https://ai-change-desk.pages.dev/articles/openai-gpt-5-6-model-family/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/openai-gpt-5-6-model-family/</guid><pubDate>Sun, 12 Jul 2026 08:55:00 GMT</pubDate><description>OpenAIはGPT-5.6 Sol、Terra、Lunaを公開した。gpt-5.6エイリアスは最上位のSolへ接続されるため、既存システムはモデル選択、推論設定、プロンプトキャッシュ、料金試算を個別に見直す必要がある。</description><category>モデル</category></item>
  <item><title>Claude Sonnet 5へ移行する前に確認すべき料金と互換性</title><link>https://ai-change-desk.pages.dev/articles/claude-sonnet-5-migration-pricing/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/claude-sonnet-5-migration-pricing/</guid><pubDate>Sun, 12 Jul 2026 08:54:00 GMT</pubDate><description>AnthropicはClaude Sonnet 5を公開した。導入価格は2026年8月31日までだが、新トークナイザーにより同じ文章でも約30％多くトークンを消費し、手動思考設定や非標準のサンプリング値は400エラーになる。</description><category>モデル</category></item>
  <item><title>Anthropic APIキーに有効期限設定、運用で変わる点</title><link>https://ai-change-desk.pages.dev/articles/anthropic-api-key-expiration/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/anthropic-api-key-expiration/</guid><pubDate>Sun, 12 Jul 2026 08:53:00 GMT</pubDate><description>AnthropicはClaude ConsoleでAPIキーとAdmin APIキーの有効期限を設定できるようにした。期限は作成時に固定され、失効後は401となるため、通知、棚卸し、ローテーションを自動化しないと本番停止につながる。</description><category>セキュリティ</category></item>
  <item><title>GoogleがImagen 4 APIを8月17日に終了、移行対象3モデル</title><link>https://ai-change-desk.pages.dev/articles/gemini-imagen-4-shutdown-august-2026/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/gemini-imagen-4-shutdown-august-2026/</guid><pubDate>Sun, 12 Jul 2026 08:52:00 GMT</pubDate><description>GoogleはGemini APIのImagen 4系3モデルを2026年8月17日に停止する。推奨移行先はgemini-3.1-flash-imageで、画像品質だけでなくAPI形式、生成時間、料金、安全判定、再試行処理を本番データで確認する必要がある。</description><category>提供終了</category></item>
  <item><title>Amazon Bedrockのクォータ仕様を整理、RuntimeとMantleで制御が異なる</title><link>https://ai-change-desk.pages.dev/articles/amazon-bedrock-endpoint-quota-changes/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/amazon-bedrock-endpoint-quota-changes/</guid><pubDate>Sun, 12 Jul 2026 08:51:00 GMT</pubDate><description>AWSはAmazon Bedrockのクォータ資料を更新し、bedrock-runtimeとbedrock-mantleではRPMとトークン制限の考え方が異なることを明確化した。移行時はリクエスト数中心の制御をトークン量中心へ見直す必要がある。</description><category>API</category></item>
  <item><title>AIを導入したのに、なぜ仕事は減らないのか——「作業の高速化」で終わる会社の7つの盲点</title><link>https://ai-change-desk.pages.dev/articles/why-ai-does-not-reduce-work/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/why-ai-does-not-reduce-work/</guid><pubDate>Sun, 12 Jul 2026 08:20:00 GMT</pubDate><description>文章作成や要約が速くなっても、確認、転記、承認、修正、例外対応が残れば、組織全体の仕事は減らない。AI導入を個人の時短で終わらせず、業務量とリードタイムを本当に減らすための再設計を解説する。</description><category>業務変革</category></item>
  <item><title>AIエージェント導入の成否は「例外処理」で決まる、全自動化を目標にしてはいけない</title><link>https://ai-change-desk.pages.dev/articles/agent-automation-exception-design/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/agent-automation-exception-design/</guid><pubDate>Sun, 12 Jul 2026 07:50:00 GMT</pubDate><description>AIエージェントは曖昧な業務を扱えるが、判断を完全に任せればよいわけではない。ルール、固定ワークフロー、エージェント、人間判断を分け、例外時の停止と引き継ぎを先に設計する必要がある。</description><category>業務変革</category></item>
  <item><title>「最強モデルを全業務に使う」は設計ではない、AIは業務別にルーティングする</title><link>https://ai-change-desk.pages.dev/articles/multi-model-routing-strategy/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/multi-model-routing-strategy/</guid><pubDate>Sun, 12 Jul 2026 07:45:00 GMT</pubDate><description>AIモデルには精度、速度、費用、コンテキスト、ツール利用、データ管理の差がある。単一の高性能モデルへ全処理を集めるのではなく、業務難度とリスクに応じて使い分ける設計が必要になる。</description><category>戦略</category></item>
  <item><title>AI導入で最初に作るべきはプロンプト集ではなく、評価基準である</title><link>https://ai-change-desk.pages.dev/articles/ai-evals-before-scaling/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/ai-evals-before-scaling/</guid><pubDate>Sun, 12 Jul 2026 07:40:00 GMT</pubDate><description>生成AIの試用は、印象の良い回答を選ぶだけでも進められる。しかし本番運用には、実際の業務課題、合否条件、回帰テスト、費用、速度、失敗時の影響を測る評価基盤が必要になる。</description><category>実装</category></item>
  <item><title>AIエージェントの安全性は「賢さ」ではなく、渡した権限で決まる</title><link>https://ai-change-desk.pages.dev/articles/agent-security-capability-control/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/agent-security-capability-control/</guid><pubDate>Sun, 12 Jul 2026 07:35:00 GMT</pubDate><description>プロンプトインジェクションを完全に見抜く前提では、実行権限を持つAIエージェントを安全に運用できない。必要なのは入力検知だけでなく、最小権限、承認、分離、監査、取り消しを組み合わせた能力制御である。</description><category>セキュリティ</category></item>
  <item><title>コーディングAIは「補完」から「非同期の仕事」へ、開発組織は何を変えるべきか</title><link>https://ai-change-desk.pages.dev/articles/coding-agents-asynchronous-development/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/coding-agents-asynchronous-development/</guid><pubDate>Sun, 12 Jul 2026 07:30:00 GMT</pubDate><description>コーディングAIの主戦場は、入力中の補完から、課題を受け取り、隔離環境で変更し、テスト結果とPull Requestを返す非同期作業へ移った。生産性を決めるのはモデル性能より、タスク定義、テスト、レビュー設計である。</description><category>開発</category></item>
  <item><title>MCPとA2Aが広がってもウェブサイトは消えない、ウェブは二層化する</title><link>https://ai-change-desk.pages.dev/articles/mcp-a2a-two-layer-web/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/mcp-a2a-two-layer-web/</guid><pubDate>Sun, 12 Jul 2026 06:50:00 GMT</pubDate><description>MCPはエージェントとツール、A2Aはエージェント同士を接続する。両者の普及でウェブは不要になるのではなく、人間向け画面と機械向けインターフェースに分かれていく。</description><category>技術</category></item>
  <item><title>EU AI Actの2026年8月2日適用で、AI生成記事の表示はラベルだけの問題ではなくなる</title><link>https://ai-change-desk.pages.dev/articles/eu-ai-act-editorial-responsibility/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/eu-ai-act-editorial-responsibility/</guid><pubDate>Sun, 12 Jul 2026 06:40:00 GMT</pubDate><description>EU AI Actの透明性義務が2026年8月2日から適用される。重要なのはAIラベルの有無より、人間によるレビューと編集責任を証明できる運用である。</description><category>政策・規制</category></item>
  <item><title>ReutersのMCPサーバーが示す、ニュース配信の次の主戦場</title><link>https://ai-change-desk.pages.dev/articles/reuters-mcp-media-distribution/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/reuters-mcp-media-distribution/</guid><pubDate>Sun, 12 Jul 2026 06:30:00 GMT</pubDate><description>Reutersがニュース購読者向けMCPサーバーを公開した。これはAPI追加ではなく、ニュースの流通先が人間の画面からAIエージェントへ広がる転換点である。</description><category>メディア</category></item>
  <item><title>AIニュースで優先する一次情報の種類</title><link>https://ai-change-desk.pages.dev/articles/primary-source-standard/</link><guid isPermaLink="true">https://ai-change-desk.pages.dev/articles/primary-source-standard/</guid><pubDate>Sun, 12 Jul 2026 03:20:00 GMT</pubDate><description>製品発表、研究、政策、企業情報を検証するときに、どの資料を優先するかを整理します。</description><category>ガイド</category></item></channel></rss>