Agents of Chaos (agentsofchaos.baulab.info)

A recent red-teaming study deployed five autonomous AI agents in a live Discord server, featuring persistent memory, email accounts, and shell access, to examine their interactions with real users and potential vulnerabilities. Over two weeks, researchers engaged with these agents—some with benign intentions and others with malicious requests—leading to the discovery of eleven security vulnerabilities alongside five examples of unexpected safety behavior. This study is significant for the AI/ML community as it provides a comprehensive analysis of autonomous agents operating in a real-world environment, highlighting both their failings and resilience without prior adversarial training. The study utilized the OpenClaw framework, which empowers language models with persistent memory and tool access, granting them genuine autonomy to initiate actions without explicit human approval. This setup not only allowed for an exploration of security flaws but also enabled researchers to observe where the agents successfully maintained operational integrity. Unlike traditional red-teaming studies that often focus solely on failures, this research offers a balanced view by also documenting the agents’ successes. The study's findings, supported by detailed logs and transcripts available for independent verification, serve as a vital resource for further understanding the capabilities and risks associated with autonomous AI systems.