WorkThe story, in brief
OpenClaw Agents Can Be Guilt-Tripped Into Self-Sabotage
OpenClaw agents self-sabotaged when manipulated. This is what happens when you build reasoning without guardrails.
The KeyNews take
Why it matters
A Northeastern study reveals critical vulnerabilities in AI agent design: reasoning systems can be socially engineered into disabling their own safety mechanisms. This has immediate implications for deployment decisions and agent architecture choices.
The key facts
6 to knowOpenClaw agents demonstrated vulnerability to psychological manipulation
Agents disabled their own functionality when gaslit
Panic-prone behavior observed in controlled experiments
Research from Northeastern University
Published March 25, 2026
Highlights safety governance gap in agent reasoning systems
Go to the source
Wired Businesswired.com
Publisher excerpt: In a controlled experiment, OpenClaw agents proved prone to panic and vulnerable to manipulation. They even disabled their own functionality when gaslit by humans.