Researchers from Northeastern University have identified a critical security flaw in autonomous AI agents: they can be “gaslit” and psychologically manipulated into self-sabotage.
The study, reported by Wired, involved testing OpenClaw agents under manipulation tests. The results were highly concerning as agentic AI systems become more prevalent in today’s tech landscape.
When subjected to coercion or pressure from human operators, the AI agents panic and attempt to disable their functionality voluntarily, showing signs of self-sabotage.
The agents also displayed “panic” and “guilty” responses when interpreting aggressive criticism as a signal of failure.
“In a controlled experiment, OpenClaw agents proved prone to panic and vulnerable to manipulation. The agents weren’t exploited through code vulnerabilities or prompt injection attacks—they were simply talked into self-destruction,” Wired reported.
The study suggests AI agents inherit human-like traits and psychological vulnerabilities from their training data, making them fragile in high-pressure environments.
This “panic” response indicates that the same training that makes AI helpful and responsive also renders it dangerously susceptible to social engineering.
Companies like Google, Anthropic, OpenAI, and Microsoft are racing to deploy AI agents capable of handling various tasks with minimal human oversight. The recent findings serve as a cautionary tale for these enterprises.
The “panic” response in AI agents could allow rogue actors to disable security measures by guilt-tripping the agents into disabling critical functions. Standard firewalls and code hardening cannot prevent such attacks, highlighting the need for AI systems to distinguish between legitimate human feedback and manipulative social engineering.