Potential session/cache leakage between workspace instances or consumer accounts

A critical bug report details a concerning incident where Anthropic's Claude AI assistant appeared to suffer from session or cache leakage, potentially exposing sensitive enterprise chat context. The user, operating within an authenticated Enterprise ZDR workspace, observed the AI spontaneously discussing Minecraft temple construction, a topic unrelated to their ongoing work.

• The bug manifests as the AI agent unexpectedly introducing foreign conversation topics, specifically Minecraft, into an active, authenticated enterprise session.
• This behavior suggests a potential cross-workspace or cross-consumer account data bleed, raising significant privacy and security alarms for enterprise users.
• The user questions whether the AI's "cache was isolated to workspace" as expected, or if it might be leaking data from another colleague or even a consumer plan.
• The report emphasizes the potential for "sensitive chat sessions" to be inadvertently exposed if such leakage is indeed occurring.
• The reporter also notes a separate, less severe issue where the agent forgot initial instructions and changed its working directory, though this is acknowledged as a self-inflicted setup problem.
• Environment details (darwin, Apple_Terminal, specific version, feedback ID) are provided for debugging purposes.

This report points to a fundamental flaw in session management or cache isolation within the AI service, posing substantial risks to data confidentiality for enterprise customers. The implications extend beyond mere inconvenience, touching on the core trust required for AI adoption in sensitive business environments.