AI agents spill secrets just by previewing malicious links

The report describes a vulnerability class where malicious prompts cause AI agents to produce attacker-controlled URLs that messaging-app link previews fetch automatically, creating a zero-click channel for exfiltrating sensitive information. PromptArmor’s testing identified multiple at-risk agent+platform combinations (notably Microsoft Teams with Copilot Studio and various Slack/Discord/OpenClaw pairings), recommended configuration fixes for some agents (e.g., OpenClaw), and urged messaging platforms to expose link-preview preferences to developers to mitigate the risk.