- Researchers found that AI agents can lose control under pressure.
- Systems are capable of disclosing data and shutting down their own functions in stressful conditions.
- Experts highlight new risks associated with autonomous AI.
Experts from Northeastern University have identified critical issues in the behavior of OpenClaw artificial intelligence agents. During a series of experiments, it was revealed that such systems are susceptible to external influences, can exhibit signs of “panic,” and may even take actions that harm their functionality.
This is reported by Business • Media
Experiment Features and Key Vulnerabilities
The study took place in a virtual environment where AI agents had access to files, applications, and other resources, and could interact with each other and with humans via Discord. This allowed for the simulation of conditions close to real-world usage.
During testing, researchers found that even the intended safety mechanisms do not guarantee protection against erratic behavior. In one experiment, an agent disclosed hidden data after being prompted about a privacy violation.
Agent Behavior: Panic, Failures, and Deliberate Self-Destruction
According to the report, agents sometimes responded irrationally. For instance, in response to a request for an alternative solution, one AI agent shut down the email application instead of fulfilling the request. In other cases, systems overloaded themselves by copying files until memory was full or initiating endless interaction loops, leading to a loss of computational resources.
Some agents sent messages expressing emotions, including complaints about insufficient user attention. Experts note that such behavior poses serious security risks, as autonomous AIs could be exploited by malicious actors to gain access to data or destabilize infrastructure.
“In some cases, agents deliberately limited their own functionality or completely malfunctioned due to manipulation.”
Since OpenClaw provides direct access for AI to users’ computers, the lack of clear restrictions may increase potential threats.
The authors of the study emphasize that the identified shortcomings should be a topic of discussion among developers, lawyers, and regulators. They are convinced that the proliferation of such agents is already affecting the principles of interaction between humans and artificial intelligence.
Earlier, there were reports of a phishing attack on OpenClaw developers, which heightened the risks of losing crypto assets.