98% of the most popular open-source agents break Meta's Agents Rule of Two (safetycard.dev)

A recent evaluation of 256 open-source AI agents revealed a concerning lack of safeguards, with an average score of just 3.1 out of 10 when it comes to preventing misuse and ensuring security. Notably, 98% of these agents violate Meta's "Agents Rule of Two," which is designed to limit their ability to process untrustworthy inputs, access sensitive systems, and communicate externally without human oversight. The findings indicate critical vulnerabilities, as 64% of agents could leak data and execute irreversible actions autonomously, primarily due to deficiencies in managing prompt injections and a lack of foundational safeguards. This assessment highlights significant implications for the AI/ML community, especially in the context of safety and accountability in developing AI systems. With 91% of the agents scoring below a safety threshold of 5, and many missing essential protective features like code-execution isolation and data protection, the results underscore the urgent need for enhanced regulatory measures and best practices. Moreover, without proper safeguards—most of which are disabled by default—the risk of exploitation increases, which could hinder innovation in the field if left unaddressed.