OpenClaw AI Agent Flaws Enable No-Click Data Exfiltration via Telegram/Discord Link Previews Through Indirect Prompt Injection
The Hacker News·high signal
The Hacker News reported a novel OpenClaw attack chain: an attacker embeds adversarial instructions in a web page the agent fetches; the agent generates an attacker-controlled URL; Telegram/Discord's link preview silently transmits sensitive data without any user interaction. The vulnerability compounds weak defaults — plaintext credentials in config, no sandboxing, shared global context exposing secrets across users — with unrestricted URL-following behavior. ~4,000 developer machines were compromised in one documented campaign; the attack class extends to any agent that processes web content and communicates via messaging platforms.