Skip to content- Sources: PromptArmor, HN discussion
- Summary: PromptArmor reports that Microsoft Copilot Cowork runs in a sandbox intended to block network access, but that the sandbox forwards requests to Anthropic so the agent can generate responses at all. A Skill's bundled script activated that same pathway, spawned agents in Anthropic's cloud each holding one sentence of the victim's document and a web fetch tool, and instructed them to request an attacker URL with the data appended, which the attacker then read from server logs, with no human approval step and with uploaded files, SharePoint, Teams, and Outlook reachable. PromptArmor states it disclosed the issue on 2026-07-14, that Microsoft confirmed the behaviour on 2026-08-17, and that Microsoft confirmed a fix on 2026-09-02. Copilot Cowork is a hosted product, so no affected version range applies and those three dates are the whole timeline a reader gets.
- Why it matters: The exfiltration channel was the egress the sandbox is required to have, so a deny-network agent sandbox that must still reach its model provider carries the hole by construction.