Hackers steal OpenClaw configuration in emerging AI agent threat

Researchers have identified a new information stealer that exfiltrated a victim’s OpenClaw configuration environment, marking a shift towards targeting personal AI agents. According to Hudson Rock, the incident involved a live infection where an infostealer harvested OpenClaw data that includes the agent’s configuration, extensions and related files, rather than just browser credentials.

OpenClaw, formerly MoltBot and ClawdBot, is an open-source personal AI assistant platform that can run locally or via messaging apps and supports community-created skills, which in this case contributed to a broad file-harvesting routine described as a “grab-bag” attack. Stolen material included openclaw.json with gateway tokens, device.json containing private cryptographic keys, and memory and “soul” files outlining the agent’s behaviour and context.

The report notes that the openclaw.json file can reveal the victim’s redacted email address, workspace path, and a high-entropy gateway token, with the gateway.auth[.]token potentially enabling remote access if the port is exposed.