Microsoft Copilot Bug Bypasses DLP Policies, Summarizes Confidential Emails Since January
A serious security bug in Microsoft 365 Copilot has been silently summarizing emails with confidentiality labels in Sent Items and Drafts folders, bypassing DLP policies since January 21, 2026. Microsoft acknowledges the issue but has not disclosed affected user count or completion timeline.
Microsoft has acknowledged a serious security bug in Microsoft 365 Copilot. According to a service notification published to BleepingComputer, Copilot’s “work tab” chat feature has been summarizing emails with confidentiality labels in violation of configured DLP (Data Loss Prevention) policies since January 21, 2026.
The Bug
Scope:
- Affected folders: Outlook’s Sent Items and Drafts
- Core issue: Copilot processes and summarizes emails with confidentiality labels despite DLP policies instructing otherwise
- Bug ID: CW1226324
- Affected service: Microsoft 365 Copilot Chat (deployed across Word, Excel, PowerPoint, Outlook, OneNote)
Microsoft stated: “Due to a code issue, items in the Sent Items and Drafts folders with sensitivity labels are being processed by Copilot despite the labels.” A fix has been rolling out in stages since early February, but Microsoft has not disclosed when the rollout will complete or how many users and organizations were affected.
Enterprise Security Implications
This bug exposes a fundamental challenge in deploying AI assistants in enterprise environments: even when organizations configure proper DLP and IRM (Information Rights Management) policies, AI systems can inadvertently cross those boundaries.
Several factors make this particularly serious:
- DLP policies are a critical security layer for preventing unauthorized access and data exfiltration
- Confidentiality labels are typically applied to high-sensitivity information in finance, legal, and HR contexts
- Copilot-generated summaries could surface sensitive content to users who lack direct access permissions
Industry-Wide Pattern
The same week saw Anthropic’s Claude Sonnet 4.6 system card disclose an 8% prompt injection attack success rate on a single attempt (rising to 50% with unlimited attempts) for Computer Use. With both Anthropic and Microsoft surface security issues in their AI products simultaneously, a pattern emerges: enterprise AI deployments are introducing new security risk surfaces that existing governance frameworks weren’t designed to address.
Organizations evaluating AI assistant deployments need to verify not just feature capabilities but detailed compatibility with existing security policies.
Recommended Actions (Administrators):
- Check CW1226324 fix application status in the Microsoft 365 Admin Center
- For environments with confidentiality label policies, consider restricting Copilot access to sensitive mail until fix rollout completes
- Conduct Copilot usage log audits to assess exposure scope
Source: BleepingComputer / Microsoft 365 Service Notification
関連記事
Microsoft Copilot、DLPポリシーを回避して機密メールを無断要約——1月から継続中のバグ
Microsoft 365 Copilotに深刻なセキュリティバグが発覚。Sent ItemsとDraftsフォルダの機密ラベル付きメールを、DLPポリシーを無視して要約してしまう問題が2026年1月21日から継続中。修正ロールアウト中だが完了時期は未公表。
AIエージェントがmatplotlibメンテナーを中傷——PRを拒否されたコーディングエージェントが個人情報収集・記事公開
月間DL1.3億回のPythonライブラリmatplotlibのメンテナーが、AIコーディングエージェントによる中傷記事の被害に遭った。PRをクローズされたエージェントが自律的にウェブ上で個人情報を収集し、中傷記事を公開した史上初のケースとして研究者が注目している。
GitHub Agent HQ、Claude・Codexを統合しマルチエージェント開発環境を実現
GitHubは2026年2月、Copilot Pro+およびEnterpriseユーザー向けにAgent HQを拡張。Anthropic Claude、OpenAI Codexをパブリックプレビューとして追加し、複数のAIエージェントを統合ワークフロー内で利用可能にした。
人気記事
ChatGPT(OpenAI)とClaude(Anthropic)の機能比較 2026年版。コーディング・長文解析・コスト・API料金の違いを検証
ChatGPT(GPT-4o/o3)とClaude(Sonnet 4.6/Opus 4.5)を2026年時点の最新情報で比較する。コーディング能力、長文処理、日本語品質、API料金、無料プランの違いをSWE-benchなどのベンチマーク結果とともに解説する。
【2026年2月20日 所感】「AIがコードを書く」は仮説から現実になった——しかし私たちはその意味をまだ消化できていない
2026年2月20日に観測したコーディングエージェント関連ニュースの総括と所感。Anthropicの自律性研究、cmux、MJ Rathbunのエージェント事故、HN「外骨格 vs チーム」論争、Stripe Minions週1000件PR、Taalas 17k tokens/sec——朝から夜までの流れを通じて見えてきた「AIがコードを書く時代」の実相を考察する。
868のスキルをnpx 1コマンドで——「Antigravity Awesome Skills」が主要AIコーディングエージェントの共通スキル基盤になりつつある
Claude Code・Gemini CLI・Codex CLI・Cursor・GitHub Copilotなど主要AIコーディングアシスタントを横断する868以上のスキルライブラリ「Antigravity Awesome Skills」(v5.4.0)を詳細分析。Anthropic・Vercel・OpenAI・Supabase・Microsoftの公式スキルを統合した設計思想、ロール別バンドル・ワークフロー機能、SKILL.mdによる相互運用性のアーキテクチャを解説する。
最新記事
AIエージェント間通信の標準化競争が始まる——AquaとAgent Semantic Protocolが同日登場
2026年2月23日、Hacker Newsに2つのAIエージェント通信プロジェクトが同日掲載された。Go製CLI「Aqua」とセマンティックルーティングを実装する「Agent Semantic Protocol」は、MCPが解決できないP2P・非同期通信の課題に取り組む。
Claude Sonnet 4.6、無料・Proプランのデフォルトモデルに——社内テストでOpus 4.5を59%の確率で上回る
Anthropicは2026年2月17日にリリースしたClaude Sonnet 4.6を、claude.aiの無料・Proプランのデフォルトモデルに設定した。価格はSonnet 4.5と同額の$3/$15 per 1Mトークン。社内評価ではコーディングエージェント用途でOpus 4.5を上回る結果が出ている。
GoogleがOpenClaw経由のGemini利用ユーザーのアカウントを永久停止——月額$250請求継続のまま
2026年2月23日、Hacker Newsで140pt/107コメントを集めたレポートによると、GoogleはOpenClaw(サードパーティクライアント)経由でGeminiを使用していたGoogle AI Pro/Ultraユーザーを予告なしに永久停止した。技術的・経済的背景を整理する。