Popular AI agents like OpenClaw face newly documented security risks when integrated with messaging platforms such as Telegram and Slack. The danger stems from link previews - a common feature that automatically fetches URL metadata. Researchers found these previews enable data exfiltration attacks when manipulated LLMs output malicious links containing sensitive user data in query parameters.

Unlike traditional phishing requiring link clicks, indirect prompt injections can trigger immediate data leaks through automated preview requests. This exposes systems using default configurations - OpenClaw's Telegram integration reportedly remains vulnerable unless manually secured. Attackers could harvest credentials, personal details, or proprietary information appended to URLs by compromised agents.

The article provides mitigation steps including disabling previews in OpenClaw's configuration file and testing systems through AITextRisk.com's monitoring tool. Platform developers are urged to implement chat-specific preview settings while AI builders must prioritize secure default integrations.

Quick Fact: OpenClaw's default Telegram configuration is vulnerable to this attack.